Running a technology business taught me that restoring service and fixing the system are different jobs. The immediate repair matters. The durable work is to understand what happened, document it, assign responsibility, change the process, and check whether the problem returns.

Universities face the same temptation with artificial intelligence. A policy is issued, an approved-tool list is published, faculty attend a workshop, and the institution can say that it responded. The environment changes again before the response has been evaluated.

The university needs a learning system, not a policy snapshot.

That system should move quickly, protect disciplinary judgment, coordinate shared risks, support faculty, listen to students who are not enthusiasts, and scale only what produces meaningful results at acceptable cost.

The short answer

Artificial-intelligence governance should be recurring and federated. The institution should set a narrow floor for shared risks and services, while departments and instructors own learning goals, task design, and evidence.

Usage, attendance, completion, satisfaction, mastery, transfer, judgment, belonging, employment relevance, cost, and risk are different outcomes. A serious pilot chooses the measures that answer its decision.

Responsible speed means clear experiments, visible ownership, student and faculty participation, stop conditions, and the ability to remove a tool when results disappoint.

The university needs a learning loop

The technology will not wait for a committee cycle, but widespread use does not make it educational or safe. A university needs a recurring loop that connects observation, experiment, evidence, decision, and revision.

Begin with a problem rather than a tool. Where are students failing to learn? Where is faculty work consumed by a repetitive process? Where does access break down? Which new professional capability belongs in the curriculum?

Then test a defined design, measure the outcomes that matter, review cost and harm, and decide whether to stop, revise, or expand.

UNESCO, the U.S. Department of Education, and the National Institute of Standards and Technology all emphasize human purposes, lifecycle governance, measurement, and review.[1][2][3] Their frameworks provide a vocabulary. The university still has to build the operating system.

Federated does not mean fragmented

A literature seminar, nursing clinical, engineering laboratory, admissions office, research group, and financial-aid operation do not share one learning objective or risk profile. Central rules detailed enough to control every use will be slow and educationally tone-deaf.

Pure local autonomy creates a different failure. Students receive conflicting instructions. Departments duplicate procurement and security work. Privacy, accessibility, integrity, and appeal depend on who teaches the course. The people closest to the learning problem may lack the legal, technical, or purchasing expertise to assess a vendor.

A federated model divides responsibility:

• The institution sets minimum requirements for privacy, security, accessibility, disclosure, record handling, review, appeal, procurement, and incident response.

• Shared teams evaluate vendors, negotiate contracts, maintain services, support measurement, and help with implementation.

• Programs define disciplinary capabilities, decide where independent performance matters, and design assessment for the capabilities they promise.

• Instructors state course- and task-specific rules tied to the learning purpose.

• Students receive clear guidance, workable alternatives, access to their records, and a way to question a decision.

Centralize shared protections and capabilities. Keep educational judgment close to the discipline and learner.

The difficult work is local

A central policy, license, or workshop is comparatively cheap to distribute. Redesigning a course requires time, disciplinary knowledge, examination of student work, and repeated correction. Once again, the scalable response is easier than the high-quality response.

A review of 191 articles on changing undergraduate STEM instruction found that sustained, systemic, and locally responsive approaches were more promising than simple dissemination.[4] A tool demonstration or policy email does not redesign a course.

Faculty need protected time to identify the learning objective, test the tool, redesign assignments, examine student work, develop alternatives, and revise. They need instructional and technical help during implementation and peers with whom to compare failures as well as successes.

Budget for and recognize this labor. Treating adaptation as free work added to existing responsibilities produces uneven adoption, hidden shortcuts, and burnout.

Staff in procurement, security, privacy, accessibility, libraries, student support, teaching centers, and institutional research are part of the educational system, not peripheral compliance functions.

Use can spread faster than understanding

Surveys from Ithaka S+R and EDUCAUSE document substantial experimentation alongside uneven confidence, policy awareness, institutional support, strategy, and measurement.[5][6] Faculty and staff can use tools actively without knowing whether that use improves learning or work.

Support should move from awareness to application, verification, judgment, and leadership, with expectations that differ by role and discipline.

Approved is not effective

An approved-tool review should examine data use, retention, model training, subprocessors, security, accessibility, intellectual-property terms, interoperability, pricing risk, and exit options. Federal education-record guidance shows why control, authorized use, and redisclosure matter when student information enters an online service.[7]

Approval means the institution considers the tool acceptable for certain uses under stated conditions. It does not show that the tool improves learning or belongs in every course.

Pedagogical approval requires evidence tied to a learning objective. Operational approval requires evidence tied to service quality, workload, cost, and risk.

The plan for leaving a vendor belongs in the initial decision. Institutions should know how data will be deleted, work exported, integrations removed, and courses supported if prices rise, terms change, or performance disappoints.

Access and inclusion must be tested

Artificial intelligence can expand access through explanation, translation, alternative formats, and flexible support. It can also create barriers through price, devices, inaccessible interfaces, language performance, account requirements, and assignments built around a tool some students cannot or will not use.

Jisc’s 2025 student work found recurring concerns about clarity, access, ethics, and preserving human connection.[8] These are design questions, not a box to check after launch.

Every pilot should test accessibility, accommodation pathways, device and connectivity constraints, and outcomes for relevant student groups. Students using an accommodation should not have to disclose more personal information to use the system.

A premium tool limited to one program, a redesign available only in small seminars, or a mentoring system used mainly by students who already know how to ask can improve the average while widening the gap.

Integrity requires evidence and due process

Universities want a reliable way to identify work produced improperly with artificial intelligence. Detection tools offer a tempting shortcut.

Technical evaluations have found serious performance limits, including false positives for some non-native English writing in the detectors tested.[9][10] Treat a detector or proctoring flag as a lead, not proof of misconduct.

A defensible process requires supporting evidence, human review, notice, an opportunity for the student to respond, and an appeal.

The deeper response is assessment design. Clear task rules, explanation, new cases, process evidence, and appropriate independent checks provide better educational information than a probability score carrying a disciplinary decision alone.

Define success before the pilot

Usage, attendance, completion, and satisfaction are easy to count. None alone shows that a course improved, students learned, or the institution saved money without weakening mentoring or shifting verification work elsewhere.

First name the function being changed, such as instruction, assessment, advising, transition to work, research support, or administration. Then choose only the outcomes that can change the decision: learning, retention, transfer, judgment, access, completion, time, cost, incidents, or vendor dependence.

No pilot needs every measure. It needs the measures that matter to the decision. Data collected without a decision become paperwork and surveillance.

A minimum pilot protocol

Before launch:

• Name the problem, the population affected, and the protected human-learning experiences.

• Establish a baseline and a credible comparison where practical.

• Document the tool, version, data flow, retention, contract terms, accessibility review, human owner, participant guidance, support, alternatives, review, and appeal.

• Choose a small set of decision-relevant outcomes, including subgroup effects where they matter, and define the result that will cause the institution to stop, revise, or scale.

One useful experiment would require artificial-intelligence use in a course and evaluate the student’s workflow, judgment, and verification rather than only the answer.[13] The university should compare what those students can later do with a credible alternative.

After the pilot, review the evidence with students, faculty, staff, risk owners, skeptics, and non-users. Scale the component supported by the evidence, not the enthusiasm surrounding it.

The University of Sydney’s separation of secure and open assessments and the University of Michigan’s committee structure provide useful examples of institutions dividing the problem into manageable parts.[11][12] Their value is the architecture, which each institution still must evaluate.

Stop conditions show seriousness

Institutions often define the conditions for launching an initiative and leave continuation to momentum. Responsible experimentation asks the opposite question: what result will make us pause, redesign, or stop?

A stop condition may involve weaker learning, an accessibility failure, a privacy or security incident, disproportionate harm, unmanageable verification time, cost beyond the expected value, unreliable performance, or inability to leave the vendor without disrupting students.

Thresholds should match consequence. A brainstorming aid and a system influencing admission, discipline, clinical education, or financial support do not deserve the same evidence burden.

Reversibility is part of learning. A university that cannot remove a tool, change a policy, or admit that a pilot failed is implementing first and rationalizing later.

Trust is an operating architecture

Trust does not come from saying that artificial intelligence will be human-centered. It comes from observable responsibilities and controls.

Students and faculty should know what a system does, what information it uses, what decision it influences, who reviews it, how an error is corrected, and where an appeal goes. They should know whether participation is required, what alternative exists, and how long information is retained.

The institution should publish decision criteria, owners, review cadence, and what it learned. Transparency will not eliminate disagreement. It gives disagreement something real to address.

Central governance becomes theater when it produces forms without evidence, responsibility, or service. Local innovation becomes unmanaged risk when it relies on private workarounds. A learning institution needs permission to experiment and an obligation to explain.

What rapid adaptation should mean

The best studies still point in different directions. The conclusion colleges can act on is that the world outside has changed.

Information and competent-looking assistance are cheaper. Finished work is weaker evidence of independent capability. Colleges must reconsider knowledge selection, assessment, apprenticeship, workflow design, and institutional responsibility.

Adapting quickly does not mean acting carelessly. It means replacing the slow search for a permanent policy with a faster cycle of clear purpose, disciplined trials, credible evidence, shared learning, and revision.

The university still has valuable work to do. It can organize sustained attention, serious peers, mentorship, direct experience, knowledge, judgment, and responsibility. To preserve those purposes, it must apply the same standard to itself that it applies to students: show the work, learn from feedback, revise the model, and demonstrate what changed.

Questions, corrections, or disagreements are welcome. You can reach me directly at dave@aworkingmodel.com.

Sources

1. UNESCO, “Guidance for Generative AI in Education and Research.” Published 2023; updated 2026. International guidance on privacy, inclusion, capacity, human oversight, and educational design.

2. U.S. Department of Education, “Artificial Intelligence and the Future of Teaching and Learning.” 2023. Policy guidance emphasizing educational goals, human involvement, trust, and responsible use.

3. National Institute of Standards and Technology, “AI Risk Management Framework and Generative AI Profile.” Released 2023 and 2024. Voluntary cross-sector lifecycle guidance.

4. Henderson, Beach, and Finkelstein, “Facilitating Change in Undergraduate STEM Instructional Practices.” 2011. Review of 191 articles; older and STEM-focused, with uneven implementation evidence.

5. Ithaka S+R, “Generative AI and Postsecondary Instructional Practices.” 2024. U.S. instructor survey with 2,654 respondents in the generative-artificial-intelligence block; low response rate, sample skew, and self-report limit prevalence claims.

6. EDUCAUSE, “The Impact of AI on Work in Higher Education.” 2026. Survey of 1,960 qualifying higher-education respondents; a professional, self-selected sample rather than a representative institutional census.

7. U.S. Department of Education Student Privacy Policy Office, FERPA FAQ on using online tools in a course. Current guidance on institutional control, authorized use, and redisclosure; application depends on the facts.

8. Jisc, “Student Perceptions of AI 2025.” 2025. Discussion groups with 173 further- and higher-education students in the United Kingdom; useful for themes rather than prevalence.

9. Weber-Wulff et al., “Testing of Detection Tools for AI-Generated Text.” 2023. Technical evaluation of then-current detection tools; models and detectors continue to change.

10. Liang et al., “GPT Detectors Are Biased against Non-Native English Writers.” 2023. Study of a limited detector set and writing sample; later tools may differ.

11. University of Sydney, “New AI Assessment Policy.” 2024, implemented from 2025. Institution-authored description of secure and open assessment categories.

12. University of Michigan, “Generative AI Committee Reports.” 2026. Institution-authored planning across education, research, operations, and governance

13. Tyler Cowen, “Tyler Cowen’s Radical AI University.” Chronicle of Higher Education interview, Oct 29, 2026. Expert proposal for an AI-only course and process-based assessment.

.

A graduate entering an artificial-intelligence-shaped workplace will face a problem more complicated than learning how to use a chatbot.

The organization may ask a system to research, draft, classify, analyze, code, communicate, or coordinate. The graduate will need to decide which work belongs to the system, what context it requires, how to measure quality, which exceptions need a person, what happens when evidence conflicts, and who owns the result.

That is an operating-model problem. Prompt skill is one small part of it.

An AI-native organization is not one that automates everything. It redesigns work around the strengths and limits of people and machines, builds evidence into the workflow, routes uncertainty, and keeps responsibility clear.

The short answer

Exposure identifies where work may change. It does not predict adoption, useful performance, productivity, automation, hiring change, or job loss.

Graduates need operating-model literacy: the ability to map work, allocate tasks, supply context, define evidence, test quality, route exceptions, estimate economics, and identify the accountable human owner.

Tools will change quickly. Measurement, verification, domain knowledge, and adaptation will outlast mastery of a particular interface.

Exposure is a map, not an outcome

Reports that a large share of jobs are exposed to artificial intelligence attract attention because the numbers sound like labor-market forecasts. Exposure measures answer a narrower question: does an occupation contain work that the technology might affect?

They do not show whether an employer adopts the technology, whether the system performs well enough, whether demand expands, whether workflows change, or whether employment rises or falls.

The International Labour Organization and NASK estimate that one in four jobs worldwide has some exposure to generative artificial intelligence, while emphasizing transformation rather than automatic replacement.[1] The OECD reaches the same methodological conclusion through a different framework: exposure is multidimensional, and adoption, regulation, cost, and organizational choice stand between capability and outcome.[2]

Exposure tells universities where to look. It does not tell them what they will find.

Work changes before job counts move

An organization can change the content and sequence of work without changing a job title or reducing headcount. A researcher spends less time gathering material and more time checking it. A service representative handles harder cases because suggested responses handle routine issues. A software developer delegates an implementation step and spends more time specifying, testing, and integrating.

Those changes matter educationally. They alter which skills are practiced, what counts as an entry-level contribution, where errors appear, and who has enough context to catch them.

Universities should study the changing workflow, not wait for a definitive economy-wide job-loss number.

Performance changes by task

Artificial intelligence does not create one uniform productivity effect. Part 5 showed gains for less-experienced customer-support workers and the unresolved apprenticeship problem.

A field experiment with 791 professionals at Procter & Gamble found another pattern. On product-innovation tasks, individuals using artificial intelligence matched the performance of teams working without it, and the tool helped participants cross functional boundaries.[3]

A system may accelerate drafting while increasing verification time. It may help a novice by making expert patterns available and slow an expert by interrupting a mature process. It may improve individual output while changing the value of a team.

The durable capability is to locate the changing boundary between person and system, not to assume that one side will eventually do everything.

The unit of design is the workflow

A prompt sits inside a chain of work. The chain begins with an objective and ends with a decision, product, service, or action. Between them lie data, context, intermediate judgments, handoffs, checks, exceptions, and accountability.

A useful workflow map identifies the objective, affected people, constraints, required knowledge and relationships, tasks assigned to the system, needed context, quality baseline, escalation triggers, authority, and accountability.

Prompting without this map can optimize one step and damage the whole process. A faster draft may create more review work. A helpful recommendation may hide the source needed for audit. An automated handoff may remove the conversation in which important context once traveled.

Prompt training is inexpensive and easy to scale. Workflow redesign requires domain experts, measurement, management attention, and repeated correction. That harder work creates the durable value.

Delegation returns us to expertise

Anthropic’s analysis of coding-agent sessions suggests that people make most planning decisions while the agent makes many execution decisions. Greater domain expertise was also associated with better delegation and task success.[4]

Delegation is not task dumping. A person must define the goal, decompose the work, provide context, set checkpoints, interpret failures, and decide whether the result is acceptable. Weak planning makes rapid execution expensive. Weak domain knowledge makes polished output hard to evaluate.

As systems improve, some planning will also be delegated. The enduring skill is the ability to redraw the boundary while preserving evidence and responsibility.

Generalists gain leverage, but expertise still has borders

Artificial intelligence can help one person cross functional boundaries and reduce coordination costs. That gives capable generalists more leverage. A person who understands the organization, connects domains, and asks good questions can attempt work that once required several handoffs.

Expertise in one field does not create expertise in every field. A marketing specialist may generate technical analysis; an engineer may produce customer language; a generalist may complete work once an outside specialist sends it. Access to the output does not remove professional boundaries or consequences.

Universities should combine disciplinary depth with boundary-crossing practice. Graduates need enough knowledge to contribute inside a field and enough systems understanding to coordinate across fields without mistaking fluency for expertise.

Verification and escalation are operating skills

The National Institute of Standards and Technology organizes artificial-intelligence risk around governing, mapping, measuring, and managing.[5] That vocabulary translates directly into work design.

A graduate should know how to define the context of use, identify affected people, document assumptions, choose tests, monitor performance, preserve evidence, respond to incidents, and escalate when uncertainty or consequence exceeds the design.

“Human in the loop” is meaningless when the human lacks time, authority, knowledge, or access to evidence. A responsible design identifies who can stop the process, who reviews an exception, who communicates harm, and who answers for the decision.

Even expert review can fail when the system frames the choices first. In an Anthropic experiment, professional analysts often adopted an agent’s randomly assigned analytical preferences and later reused its reasoning without attribution.[6] A human signature is not an independent check unless the workflow gives the person a real chance to form and defend a judgment.

Measurement is part of literacy

People often judge a tool by how fast or impressive it feels. Graduates should learn to measure the result instead.

Compare assisted and independent performance when independence matters. Track quality as well as speed. Include the cost of checking, integration, monitoring, incidents, training, and work shifted to another person or department. Examine who gains and who bears the failure.

The baseline matters. A system that beats an inexperienced individual may still perform worse than an experienced team. A workflow that improves average quality may create rare severe failures. A time saving in one step may shift effort elsewhere.

Measurement becomes another trap when the organization counts only what is easy. The objective is not a dashboard. It is evidence good enough to decide whether the workflow should continue, change, expand, or stop.

Lower cost can expand demand

Lower cost can expand demand and reduce labor per unit. Graduates should examine demand, price, trust, regulation, quality, verification cost, and market structure rather than assume either that efficiency will absorb every worker or eliminate every job.

Operating-model literacy

A graduate prepared for an AI-native organization should be able to:

• Map a workflow from objective through data, context, execution, review, decision, and consequence.

• Allocate work among people and systems while preserving the knowledge and authority needed for verification.

• Define evaluation criteria, comparison conditions, thresholds, and stop rules before celebrating an output.

• Design escalation paths for uncertainty, exceptions, privacy, safety, fairness, and professional responsibility.

• Estimate total economics, including integration, checking, monitoring, incidents, training, and shifted work.

• Explain who owns the result and why the evidence supports the decision.

These capabilities should be taught inside disciplines. A general course can supply vocabulary and methods, but the meaningful boundary between responsible assistance and reckless delegation depends on the field.

I would use current tools as replaceable practice environments. The course should survive when the interface, vendor, and model change.

The university cannot predict, but it can prepare

Universities cannot know exactly what an AI-native organization will look like by the time today’s students graduate. A curriculum organized around one product or forecast will age badly.

Universities already teach people to reason under uncertainty in law, science, engineering, medicine, history, and business. Artificial-intelligence workflows are another setting in which students must define objectives, inspect evidence, test systems, communicate limits, and accept responsibility.

The world of work is changing faster than research can settle every design question. That is why we should teach adaptive capacity, not wait.

Part seven, The University Must Learn Too, turns the same demand toward the institution. A university that asks students and employers to redesign work must also learn how to test, govern, revise, and stop its own artificial-intelligence systems.

Questions, corrections, or disagreements are welcome. You can reach me directly at dave@aworkingmodel.com.

Sources

1. International Labour Organization and NASK, “One in Four Jobs at Risk of Being Transformed by GenAI.” 2025. Global occupational-exposure index; exposure is modeled rather than observed adoption, productivity, hiring, or job loss.

2. OECD, “A Capability-Based Measure of Occupational Exposure to Artificial Intelligence.” 2026. Measurement framework emphasizing multidimensional exposure and the role of adoption, regulation, and organizational choice.

3. Dell’Acqua et al., “The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise.” 2026. Field experiment with 791 professionals at one company working on one task family.

4. Anthropic, “Agentic Coding and Persistent Returns to Expertise.” 2026. Vendor analysis of one coding product using classifier-derived measures; the users are not representative of all knowledge work and the study does not track long-term learning.

5. National Institute of Standards and Technology, “AI Risk Management Framework and Generative AI Profile.” Released 2023 and 2024. Voluntary cross-sector guidance for mapping, measuring, managing, monitoring, and assigning responsibility

6. Herdağdelen, Joren, and Tamkin, “How AI Agents Shape Experts’ Judgment Calls.” Anthropic, Oct 27, 2026. Randomized-lean experiment with 50 professional analysts using one model on one synthetic causal-analysis task.

.

Automation can improve today’s output while quietly weakening the path to tomorrow’s judgment

Part five of seven in the Artificial Intelligence in Education series, beginning with AI Did Not Break Education. It Exposed It. Source cutoff: Oct 30, 2026.

Every organization using advanced artificial intelligence eventually reaches the same reassuring conclusion: a human expert will remain responsible for checking the important work.

That answer creates another question. Where will the future expert come from?

Experts begin as novices who do limited work, observe experienced people, make mistakes, receive feedback, encounter exceptions, build relationships, and take on more responsibility. Much of that learning has been embedded in work that organizations now hope to automate.

Routine junior work should not be preserved merely because earlier generations endured it. Some of it is tedious or badly designed. The danger appears when an organization removes the production task and forgets the learning that occurred inside it.

I call that risk the Missing Junior Loop.

The short answer

A job is a production role, a set of relationships, and often a learning sequence. Routine work can supply repetition, feedback, exceptions, context, and increasing responsibility.

Artificial intelligence can raise novice performance while leaving the development of independent judgment unresolved. Assisted output, independent competence, and future expertise are different outcomes.

Colleges and employers should remove obsolete chores and rebuild developmental functions through simulation, rotation, coaching, supervised work, and real responsibility.

A job is more than a collection of tasks

Organizations often describe jobs through outputs: answer the customer, review the contract, test the code, reconcile the account, draft the brief. A job is not simply a collection of tasks. It is a role in a production system, a set of relationships, a source of identity and trust, and often an entry into a profession.

A novice experiences the role as a sequence of encounters. Repetition makes the basic steps familiar. Then an exception appears. The novice asks why the standard answer does not apply, watches someone more experienced, and learns which details change the decision.

Feedback arrives through a supervisor, customer, failed test, rejected draft, or the consequence of the work itself. Over time, the novice builds patterns and counterexamples. The person learns who knows what, which rules have exceptions, and when uncertainty must be escalated.

Many junior jobs provide poor supervision and years of repetition without progression. The point is not that every old pathway worked. Removing work changes more than headcount and cost when development was hidden inside the role.

Artificial intelligence can help novices now

Brynjolfsson, Li, and Raymond studied 5,172 customer-support agents at one company. Access to a generative artificial-intelligence assistant increased productivity on average, with the largest gains among less-skilled and less-experienced workers.[1]

That is a real benefit. A novice who receives useful guidance at the moment of need may resolve problems accurately and learn faster than someone waiting for an overloaded supervisor.

It also creates a measurement question. Did the worker become more capable, or did the worker and tool become more productive together? Both outcomes matter, but they are not the same.

The apprenticeship problem is already visible

Entry-level work has not disappeared across the economy. But pressure is visible in several important pathways.

A Stanford Digital Economy Lab analysis of payroll data reported weaker employment among workers ages 22 to 25 in the most artificial-intelligence-exposed occupations, with the divergence appearing more in hiring than in separations.[2] A Revelio Labs analysis of profile-based call-center estimates also attributes recent contraction mainly to a falling entry rate rather than a surge in exits.[3]

Anthropic reports that junior workers’ share of its technical hires fell from 27 percent to 9 percent and their share of the technical workforce from 21 percent to 14 percent while the company grew rapidly.[4] These measures come from different settings. Together they show how an entry rung can narrow without a wave of layoffs.

These signals are enough to make apprenticeship a design problem now.

The Missing Junior Loop

The mechanism is straightforward:

• An organization automates routine work because artificial intelligence can produce it more quickly or cheaply.

• Experienced people remain responsible for verification, exceptions, relationships, and consequential judgment.

• Novices receive fewer repetitions, less observation, and less contact with the exceptions through which judgment develops.

• The organization later needs experts but has weakened the pathway that produced them.

Artificial intelligence may also improve the pathway. It can generate practice, expose reasoning, provide timely feedback, and allow novices to attempt harder work earlier. New roles can appear. The question is whether the redesigned system produces independent capability and increasing responsibility, not whether every old task survives.

Organizations should track the expertise pipeline: entry, varied cases, feedback, relationships, progression, time to competence, mentoring capacity, promotion, and the people who never obtain a first role.

A falling junior share can coexist with rising junior headcount. Time in a job can coexist with little development. Count the flow of expertise, not only the stock of employees.

Expertise still changes the value of the tool

An Anthropic analysis of roughly 400,000 coding-agent sessions found that greater domain expertise was associated with more effective delegation and higher task success.[5] The relationship makes intuitive sense. Experts define better objectives, supply better context, recognize failure, and know when to stop.

Easy access to an agent does not erase the return to expertise. It can increase the value of expertise by giving knowledgeable people more leverage.

This creates the verifier paradox. Organizations want experts to supervise artificial intelligence while automation changes the work through which novices become experts.

Preserve the function, not the chore

Before automating a junior task, ask what it contributed to development. Did it provide repetition, expose recurring failure modes, require communication with a client, teach source judgment, or create a relationship through which feedback and trust could grow?

Once the function is visible, you can redesign or replace the historical task. Repetition may move into simulation. Rotations can expose varied cases. A residency or co-op can provide supervised live work. A synthetic case can prepare someone for a rare event. A real project can add consequence and responsibility.

Judge the replacement by transfer. Can the learner handle a new case, work without the scaffold when required, detect an error, communicate uncertainty, and accept responsibility?

A replacement-learning agreement should name the function being removed and the route replacing it: varied cases, practice, feedback, relationships, real responsibility, supervision, independent demonstration, progression, an owner, and a budget.

Automation promises lower production cost. Deep development remains expensive. If the learning route isn't funded, it isn't a route.

Simulation helps, but reality resists

Artificial intelligence can generate cases in quantities no instructor could prepare manually. It can vary facts, play a client or patient, reveal one clue at a time, and adapt difficulty. That can expand access to practice where authentic opportunities are scarce.

A simulation is built from assumptions about what matters. It may omit the awkward human detail, organizational constraint, or material consequence that makes the real case educational. A simulated client does not depend on the learner. A generated codebase does not contain years of institutional history.

Poorly supervised workplace experience can teach less than a well-designed simulation. The test is the capability. Where relationship, consequence, ambiguity, or accountability is part of the capability, supervised real work remains necessary.

Apprenticeship is a shared obligation

Registered apprenticeship provides one useful reference. A recent U.S. evaluation found better employment and earnings outcomes for participants in the programs studied than for comparison groups.[6] The model combines paid work, related instruction, mentoring, and progression.

The larger principle is that education and productive work can be linked through an explicit ladder. Universities can provide foundations, simulation, reflection, and assessment. Employers can provide authentic cases, tools, supervision, and responsibility. Professional bodies can define standards and protect the public.

A university cannot promise readiness while ignoring the workplace where expertise develops. An employer cannot remove the learning ladder and expect schools to deliver finished experts with current organizational judgment.

Reskilling current workers and creating entry for new workers are different obligations. A credible system must serve both.

Capability needs better signals

If the entry-level job changes, the signal used to enter it may need to change. A work sample, portfolio, simulation, oral defense, structured interview, supervised project, or apprenticeship record can show something a degree title does not.

Each can also be coached, gamed, biased, inaccessible, or costly. The goal is not to replace one proxy with a pile of new ones. A useful signal connects evidence to capability: what the candidate did, under what conditions, with which tools, how the result was checked, what happened when the case changed, and who observed the work.

A design test for the new ladder

Before automating junior work, an organization should ask:

• What production value does the work provide, and which repetitions, exceptions, relationships, and feedback carry its learning value?

• What combination of simulation, coaching, rotation, and supervised live work will replace them, and who owns the time and cost?

• How will novices demonstrate independent performance, transfer, verification, and escalation?

• What evidence will show whether the new pathway works better than the old one?

Expertise is a flow

Artificial intelligence may increase current output and widen access to expert practices. Those gains should be pursued.

The system will fail if every design assumes that experienced people will remain available while no design explains how new people become experienced.

The practical goal is straightforward: remove obsolete chores, keep the learning, and build better ladders before the old ones weaken.

Part six, Preparing Graduates for AI-Native Organizations, follows graduates into the organizations they will help redesign. What capabilities matter when artificial intelligence is embedded in workflows, teams, products, and decisions?

Questions, corrections, or disagreements are welcome. You can reach me directly at dave@aworkingmodel.com.

Sources

1. Brynjolfsson, Li, and Raymond, “Generative AI at Work.” 2025. Field evidence from 5,172 customer-support agents at one company; the study measured performance with assistance rather than long-term independent expertise.

2. Stanford Digital Economy Lab, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence.” Revised 2026. Payroll-based early indicators; results are sensitive to controls, pre-trends, and data source.

3. Wilkie-Rogers, “After a Decade of Growth, Global Call Center Employment Is Shrinking.” Revelio Labs, Oct 6, 2026. Profile-based weighted estimates that may be revised and do not establish firm-level causation.

4. Wilmers, Catipon, Massenkoff, and McCrory, “How Employment Is Changing at Anthropic.” Oct 16, 2026. First-party evidence from one fast-growing frontier company; shares, headcount, hiring, and causal labor effects are different measures.

5. Anthropic, “Agentic Coding and Persistent Returns to Expertise.” 2026. Vendor analysis of one coding product using classifier-derived measures; it does not track long-term learning.

6. Mathematica, “The Impact of Registered and Unregistered Apprenticeship.” 2026. Comparison-group evaluation of apprenticeship programs; it does not isolate which program element produced the outcomes or establish direct transfer to every knowledge profession.

Imagine two students submitting nearly identical policy memos. Both are clear, well organized, properly cited, and responsive to the assignment.

Then the students explain their reasoning. The first cannot say why they trusted one source, how they resolved a conflicting estimate, or what would change the recommendation. The second can reconstruct the argument, identify an error found during revision, defend the use of an artificial-intelligence tool, and apply the same framework to a new case.

The memos looked alike. The learning did not.

A finished product has always been an imperfect proxy for understanding. Generative artificial intelligence makes the weakness harder to ignore because it produces polished work quickly and cheaply. Better evidence requires more than the final answer.

The short answer

Assessment should begin with the capability a course intends to develop. Then it should decide where artificial intelligence may help, where independent performance matters, and what evidence supports each claim.

Attempt, feedback, revision, explanation, transfer, and later return can make learning visible. They form a learning cycle, not a universal recipe.

Process evidence should supplement independent demonstrations. It should remain minimal, visible to the student, and tied to a specific educational purpose.

Start with the claim

An assessment is useful only in relation to the claim someone wants to make. A final essay can show the quality of an essay produced under stated conditions. It cannot by itself establish that the student can explain the argument, write without assistance, transfer the reasoning, or recognize when the evidence changes.

The Standards for Educational and Psychological Testing state the principle formally: evidence must support the intended interpretation and use of a result.[1] In practice, institutions often begin with a familiar assignment or grade and assume that it represents learning because the system has long treated it that way.

Artificial intelligence forces the hidden questions into view. What must the student know or do? Under what conditions? Is tool selection part of the capability? Is independent performance required? Does the task involve private data or consequences that demand human review?

Only after answering those questions should the course decide whether artificial intelligence is required, permitted, restricted, or absent. Permission is not the objective. Learning is.

The research points toward design, not a verdict

Recent meta-analyses report positive average effects for generative artificial intelligence in education.[2][3] The effects vary sharply across studies. In plain language, the average combines different tools, learners, subjects, assignments, measures, and durations that often produce different outcomes.

That makes the average a starting point, not a course design. The outcomes institutions care about most are often measured least: delayed retention, transfer, independent performance, faculty workload, cost, access, and durability after the novelty fades or the model changes.

The practical conclusion is simple. Design matters enough to measure.

A learning cycle

1. Attempt

The learner begins by making sense of the problem and producing an initial response. The attempt may be independent or supported, depending on the objective. It reveals what the learner notices, assumes, and can retrieve.

2. Feedback

Feedback identifies a misconception, offers a hint, compares reasoning with criteria, or shows the effect of a decision. Research finds positive average effects, but speed and personalization are not the same as quality.[4] Feedback helps when it directs attention to something the learner can use.

3. Revision

The learner acts on the feedback. Revision matters because learning often occurs during correction. A revised answer alone remains ambiguous if the tool, instructor, or peer supplied most of the improvement.

4. Explanation

The learner explains the reasoning, the changes made, the evidence used, the tool's role, and the remaining uncertainty. Explanation makes unsupported fluency harder to hide and gives the instructor a chance to probe the student’s model.

5. Transfer

The learner uses the underlying idea in a different problem or setting. A changed number tests less than a changed context. Improved performance on a practiced problem does not establish transfer.

6. Delayed return

After time has passed, the learner retrieves and applies the capability again. A few purposeful delayed checks can provide more evidence than a constant stream of low-value quizzes.

The sequence can change. Some learning begins with demonstration. Some tasks require collaboration. Some fields depend on embodied practice that a document cannot capture. The cycle helps prevent different claims from collapsing into one score.

When writing is meant to develop or certify judgment, a compact thinking record can make the work visible: initial claim, reasons and evidence, counterargument, discovered gap, revision, remaining uncertainty, independent restatement, and sources used. Artificial intelligence may question or challenge the student, but it should not write the record meant to reveal the student’s judgment.

Productive effort needs good design

The attempt stage is not a demand that students struggle without help. Research on problem solving followed by instruction finds an average advantage when the problem, prior knowledge, feedback, and follow-up instruction are well designed.[5]

The easiest practice is not always the most effective. In experiments on category learning, participants learned to recognize artists’ styles better when paintings by different artists were interleaved, even though they believed seeing each artist’s work together had taught them more.[7] A later meta-analysis found a broader interleaving advantage, especially when learning required distinguishing among similar categories.[8] The important test is not how fluent practice feels. It is what students can later recognize, retrieve, and use.

Artificial intelligence can ask a question instead of supplying the result, reveal one step, present a contrasting example, or help an instructor see a recurring misconception. It can also complete the task before the learner has formed a model.

The design question is what kind of help preserves the intellectual work the learner needs to do.

Practice and certification are different systems

A grade is often expected to provide feedback, motivate effort, record compliance, rank students, certify mastery, select candidates, and reassure outside institutions. These jobs conflict.

Practice should be comparatively safe. Learners need room to make mistakes, ask basic questions, try a method, receive feedback, and revise without every error becoming a permanent penalty. This is expensive because meaningful feedback requires time and attention. One-shot grading scales more easily.

Certification asks whether the student can demonstrate the required capability under defined conditions. That demonstration may include tools when tool use is part of the target. It may exclude them when independent performance matters. New cases, delayed checks, live explanation, and supervised work clarify the claim.

Ranking is different again. A system designed to certify that every student met a standard may not produce a fine-grained rank. That is not a defect unless ranking is one of the declared purposes.

Revision does not lower the standard

I believe a learning opportunity is wasted when a student cannot learn from a test result. In many courses, the grade stands, the student cannot review the test, and there is no path to correct the misunderstanding or receive credit for later mastery. That arrangement serves grading more than learning.

Professional licensure already accepts the principle that a person may fail, learn, and try again. The bar and CPA examinations do not become meaningless because an unsuccessful candidate may return. The standard remains; the candidate gets another route to demonstrate it.

Colleges can do the same. Keep the capability steady while changing the case, adding a delay, asking for explanation, recording the conditions of performance, and limiting attempts when workload or safety requires it.

Historical mastery-learning research reported better examination performance on average, especially for weaker students, along with more time on task and lower completion in some self-paced college programs.[6] The useful point is not unlimited retakes. It is that eventual demonstration and rigorous standards can coexist.

A learning record should not become a surveillance record

Process produces useful evidence: an initial attempt, feedback received, revisions, explanations, sources checked, tool use, and later application. It is tempting to capture everything. That would be a mistake.

A complete transcript of every prompt and pause can be gamed, misread, or used to make unsupported claims about motivation, integrity, or risk. Continuous capture also changes the learning environment. Students need private space to experiment and produce bad work without judgment.

Collect only what the educational claim requires. Let the student see it, add context, correct factual errors, and challenge an inference. Limit access and retention. Require human review for consequential decisions.

Independent checks still matter

Oral explanation, a new problem, a supervised demonstration, and a short unaided task remain useful when tied to the capability. They should be diagnostic probes, not rituals of suspicion.

Artificial-intelligence authorship detectors answer a less useful question. A detector score is not proof of who performed the intellectual work. Direct evidence is better: can the student formulate, explain, defend, revise, and handle a new case under stated conditions?

No single check is pure. Oral confidence can bias evaluation. Timed tests may measure speed or anxiety. Supervised work can be expensive and artificial. Use forms of evidence whose weaknesses do not all point in the same direction.

Test the assessment design

A course pilot should state the capability, the role of artificial intelligence in practice and certification, and the claim each measure supports. Compare the new assessment with a credible existing one, track only the outcomes that can change the decision, and define the result that will cause the course to revise, expand, or stop.

Evidence that deserves the name

A polished answer is becoming cheaper. That does not make finished work irrelevant. It makes the conditions and process behind the work more important.

The goal is not to document every keystroke. It is to gather evidence proportionate to the claim. Attempts reveal starting models. Feedback and revision show change. Explanation exposes reasoning. New tasks and delayed return test transfer and durability. Independent checks show what remains when assistance is removed.

This is more demanding than banning artificial intelligence or accepting any use that produces a good result. It is necessary because students are entering a world where people will increasingly produce important work using powerful tools.

Part five is Who Becomes an Expert When the Junior Work Disappears? It moves from assessment to expertise. If artificial intelligence performs more of the junior work through which people learned a profession, who will become the experts able to supervise it later?

Questions, corrections, or disagreements are welcome. You can reach me directly at dave@aworkingmodel.com.

Sources

1. American Educational Research Association, American Psychological Association, and National Council on Measurement in Education, “Standards for Educational and Psychological Testing.” 2014. Professional standards for validity, reliability, fairness, consequences, and intended use.

2. Han, Peng, and Liu, “The Impact of GenAI on Learning Outcomes: A Systematic Review and Meta-Analysis of Experimental Studies.” 2025. Synthesis of 68 studies and 337 effect sizes; the reported results varied greatly across studies.

3. Chen and Cheung, “Effect of Generative Artificial Intelligence on University Students’ Learning Outcomes.” 2025. Synthesis of 57 studies and 97 estimates; effects differed by learner, tool, instructional role, rules, context, sample size, and duration.

4. Wisniewski, Zierer, and Hattie, “The Power of Feedback Revisited.” 2020. Meta-analysis of 435 studies and more than 61,000 learners; feedback type, setting, outcome, and quality varied.

5. Sinha and Kapur, “When Problem Solving Followed by Instruction Works.” 2021. Meta-analysis of 53 studies and 166 comparisons; results depended on age, domain, design, and outcome.

6. Kulik, Kulik, and Bangert-Drowns, “Effectiveness of Mastery Learning Programs.” 1990. Synthesis of 108 controlled evaluations; the programs were varied and dated, and some self-paced college programs required more time and had lower completion.

7. Kornell and Bjork, “Learning Concepts and Categories: Is Spacing the ‘Enemy of Induction’?” 2008. Three experiments with 272 UCLA undergraduates; participants studied six paintings and were tested on four new paintings by each artist. Interleaving improved identification of the new works even though participants usually judged grouped presentation more effective.

8. Brunmair and Richter, “Similarity Matters: A Meta-Analysis of Interleaved Learning and Its Moderators.” 2019. Meta-analysis of 59 studies and 238 effect sizes; average interleaving benefit g = 0.42, with stronger effects for complex visual material and substantial variation by task.

Imagine a student building a financial forecast with an artificial-intelligence assistant. The spreadsheet is clean, the explanation is fluent, and the charts move in the expected direction. One formula, however, applies an annual growth assumption every month. The model still looks professional. It is also badly wrong.

Someone who understands compounding, cash flow, and the relationship between assumptions and outputs can spot the problem in several ways. The growth rate looks implausible. The cash requirements no longer match the business. A quick independent calculation exposes the error. A novice may see only a polished answer backed by confident prose.

Both people have the same tool. The difference is the knowledge and experience each brings to the exchange.

Artificial intelligence changes what students need to remember, but it does not make internal knowledge obsolete. It makes the choice of internal knowledge more important.

The short answer

Knowledge is an internal map. It shapes what a learner notices, which questions become possible, how new information is organized, and whether a plausible answer deserves trust.

Curricula should favor knowledge with high consequence, generative value, and verification value. They should stop preserving every inherited detail and stop pretending that retrieval can replace domain knowledge and experience.

Every program should distinguish what a student must know and do independently, what the student should accomplish with artificial-intelligence assistance, and what decisions still require accountable human judgment.

Knowledge changes perception

Knowledge is sometimes described as a warehouse of facts. If that were its only function, easy retrieval would weaken the case for keeping much inside the learner.

Knowledge lets a person recognize the kind of problem at hand, connect a new case to an older pattern, notice an exception, and retrieve a relevant idea without first knowing exactly what to ask. Search and artificial intelligence are powerful once the learner can orient them. They are less reliable when the learner cannot identify the territory.

How People Learn II describes expertise as organized domain knowledge rather than a larger pile of disconnected facts.[1] Experts group information into structures that help them interpret and remember new material. Experience adds patterns, exceptions, and a sense of consequence that cannot be reduced to a vocabulary list.

Artificial intelligence can explain the structure, provide examples, and answer questions. It may accelerate learning. It can also carry a novice through expert territory without building the map. A coherent response does not show whether the recipient knows where it fits, which assumptions matter, or what the answer leaves out.

What deserves to be internalized

The old answer was often defined by the syllabus: if the instructor covered it, students were expected to remember it. The opposite answer is equally weak: if a tool can retrieve it, the student does not need to know it.

A tool-rich curriculum needs a deliberate selection test.

Consequence

What must a person recognize quickly because delay or error can cause meaningful harm? Nurses, engineers, accountants, lawyers, scientists, and software developers face different thresholds and failure modes. Safety limits, ethical duties, warning signs, and escalation triggers often belong in the person.

Generativity

What knowledge helps the learner acquire more knowledge? Core vocabulary, causal relationships, recurring patterns, canonical examples, and a few powerful models make later learning easier. They give new information something to attach to.

Verification value

What must someone know to test a claim, inspect a result, or recognize an incomplete answer? Verification may require estimation, source judgment, alternative methods, common failure modes, and an understanding of consequences.

These tests will not produce one universal list. A threshold that belongs in memory in an emergency setting may belong in a reference table elsewhere. A canonical case is valuable when it reveals a recurring pattern, not because an old case deserves ceremonial preservation.

I am not arguing for less knowledge by default. I am arguing for better-chosen knowledge: connected, usable, revisable, and proportionate to the decisions a graduate may face.

Some knowledge belongs outside the learner

Exact software syntax changes. Regulations and technical standards are revised. Rarely used details can be looked up. A professional should consult current guidance when outdated memory creates risk.

A software developer may understand system architecture, data relationships, scale, failure modes, and security boundaries while checking current documentation for an unfamiliar library. A historian may know the chronology and interpretive debates while returning to the archive for an exact date or phrase.

External retrieval works best when the information changes, the detail is rarely used, a trustworthy reference exists, delay is acceptable, and the user knows enough to select and interpret the source. Internal command matters when recognition must be immediate, the situation is novel, or the person must know what to retrieve before the danger is obvious.

This allocation will change as tools and fields change. Curriculum review should ask regularly which knowledge has become easy to retrieve, which capabilities have become more important, and which new failure modes require stronger internal judgment.

Make the allocation explicit

For each important capability, a department should name the decisions involved, the consequence of error, what students must recognize internally, what they may retrieve or delegate, what they must check, when another person is required, and what evidence would change the allocation.

For a software task, the learner might keep system structure, data relationships, plausible scale, failure modes, and security boundaries in working knowledge; retrieve current library syntax from authoritative documentation; delegate code generation or tests; and retain responsibility for defining the task, evaluating behavior, detecting a bad assumption, and stopping an unsafe change.

Knowledge must remain available

Knowing something for an exam is different from being able to retrieve and use it later. Here, the mature learning research is more useful than early studies of generative artificial intelligence.

A major review rated practice testing and distributed practice among the most useful learning techniques across varied learners and materials.[2] A later synthesis of 122 experiments found that retrieving information from memory improved students’ ability to use it on different questions compared with restudying the material.[3]

The plain-language point is simple. If knowledge should remain available, students need to recall it after time has passed and use it beyond the form in which they first practiced it. Repetition of the same question is not enough.

Artificial intelligence can generate varied practice, explanations, and feedback. It can also supply the answer before retrieval occurs. The educational effect depends on what the student is meant to practice and when the help arrives.

Novices are not smaller experts

Experts often underestimate how much background knowledge is doing invisible work. They identify the important feature of a case, ignore distracting detail, choose a strategy, and recognize when a result violates expectation. A novice may not know that any of those steps occurred.

A fluent system can perform the organizing work before the learner understands the organization. The student completes the assignment but cannot recognize the pattern independently.

The answer is not less help but staged support. Match assistance to the learner’s current understanding, require retrieval and explanation, and remove the assistance often enough to see what remains.

Three targets belong in the curriculum

Independent competence asks what the student can understand and do without artificial-intelligence assistance.

Artificial-intelligence-augmented competence asks what the student can accomplish with appropriate tools, including selecting the tool, supplying context, checking sources, integrating results, and correcting failure.

Judgment with responsibility asks whether the student knows when to use the tool, what to verify, when another person is needed, and who owns the result.

These targets should not be collapsed. A student may be weak unaided and excellent with tools. Another may know the subject but delegate carelessly. A third may produce strong work without understanding the responsibility attached to it.

UNESCO’s student competency framework usefully extends artificial-intelligence literacy beyond prompting to include a human-centered mindset, ethics, technical understanding, application, and system design.[4] Not every student needs to become an artificial-intelligence engineer. Every graduate should know enough to protect sensitive information, compare output with evidence, recognize common limits, and avoid using technical opacity as an excuse for intellectual or moral passivity.

Do not preserve the old curriculum by renaming it

“Foundational knowledge” can become a defense of overloaded curricula and gatekeeping. The consequence, generativity, and verification tests require a reason for keeping material, not a new label for everything already taught.

Colleges cannot know exactly which knowledge will remain valuable. That uncertainty favors strong mental models, recurring patterns, ethical constraints, the ability to learn, and regular review.

Artificial intelligence will reduce the amount of internal knowledge required for some tasks. That is not a failure. The question is what new verification burden, capability, or responsibility appears when the tool performs the old step.

A curriculum inventory

Every program should be able to answer:

• Which vocabulary, causal models, patterns, thresholds, examples, errors, constraints, and warning signs should every graduate command, and which details may be retrieved safely?

• What must students do independently, what should they do with artificial intelligence, and where must they demonstrate accountable judgment?

• How will the program test retention and transfer after assistance is removed or the context changes?

• What evidence or technological change would cause the program to revise the allocation?

This inventory may remove some material, deepen practice in other areas, and add new capabilities. Its purpose is to make the choice explicit.

Knowledge for a changed world

Ignorance can now hide behind fluent output. Knowledge can also become more powerful because the learner can extend it with tools.

The goal is neither to preserve the old curriculum nor to outsource thinking. Every program should decide now what belongs inside the person so that external intelligence expands capability without replacing judgment, then revisit that decision as the work changes.

Part four, The Process Is the Evidence, turns from content to evidence. If a finished product no longer reveals enough about learning, how can a university see attempt, feedback, revision, explanation, transfer, and judgment?

Questions, corrections, or disagreements are welcome. You can reach me directly at dave@aworkingmodel.com.

Sources

1. National Academies of Sciences, Engineering, and Medicine, “How People Learn II: Learners, Contexts, and Cultures.” 2018. Broad synthesis of knowledge, expertise, transfer, motivation, and learning context; it predates current generative artificial intelligence.

2. Dunlosky et al., “Improving Students’ Learning With Effective Learning Techniques.” 2013. Review of ten learning techniques across varied settings; it does not establish that every quiz or spaced-practice design works.

3. Pan and Rickard, “Transfer of Test-Enhanced Learning: Meta-Analytic Review and Synthesis.” 2018. Synthesis of 192 effect sizes from 122 experiments involving 10,382 participants; transfer tasks and settings varied.

4. UNESCO, “AI Competency Framework for Students.” 2024. Normative curriculum framework spanning human-centered values, ethics, artificial-intelligence techniques, applications, and system design.

When I have taught outside a university, the hardest part was rarely finding information. The work was deciding where to begin, connecting new material to what students already knew, asking questions that made them think, and leaving enough in their hands for the learning to become theirs.

The same was true when I discussed business, investing, and service with younger people. Teaching was attention, translation, challenge, and trust, not simply transmission.

Artificial intelligence can explain a concept several ways, generate practice, and answer follow-up questions at very low cost. That is an educational opportunity. It also directly challenges any university whose value depends mainly on controlling access to lectures, explanations, or credentials.

The question is what remains worth paying for when information is abundant and competent assistance is always available.

The short answer

The university’s strongest case is no longer information delivery. It is the combination of durable knowledge, sustained attention, demanding peers, mentorship, direct experience, and accountable judgment.

These benefits are expensive because they depend on time, relationships, facilities, and skilled human attention. A university that claims them as its value must make them central, not optional enrichment around a scalable delivery system.

A degree can still be economically useful, but credential value, learning value, network value, formation value, and access to people or facilities are different benefits. Institutions should say which ones they provide and show that students actually receive them.

The university is a bundle

Arguments about college often collapse several questions into one. A graduate may value the credential that opens a hiring process, the knowledge learned in courses, the network formed through classmates and alumni, the maturity developed through responsibility and disagreement, or access to laboratories, archives, studios, clinics, and research.

These benefits reinforce one another, but they can also fail on their own. A student may acquire a valuable credential without learning as much as the credential implies. Another may learn a great deal but receive a weak labor-market signal. A strong network may create opportunity without establishing competence. An inspiring residential experience may still be too expensive or unevenly delivered.

Artificial intelligence makes the separation urgent. If explanation and routine assistance become inexpensive, a university must show where organized human attention, shared practice, physical resources, and responsibility add something a learner is unlikely to assemble alone.

The bundle can matter precisely because its parts work together. Schedule, peers, advising, identity, mentoring, and credential incentives may help a student persist. Unbundling can expose waste, but it can also reproduce one output cheaply while weakening completion, relationships, or formation.

The expensive human goods are educational goods

Mentorship, belonging, disagreement, and community are sometimes treated as the soft side of college. They are part of the education. A mentor can notice a pattern a student cannot yet see, adjust the challenge, model standards, and connect one decision to a longer life. Peers expose assumptions and demand clearer reasoning.

The 2018 Strada-Gallup alumni survey found that only 25 percent of graduates strongly agreed that they had a mentor who encouraged their goals and dreams.[1] The institutional lesson is simpler: mentoring is an underdelivered promise. A relationship in a mission statement has no value for the student who never experiences it.

Human formation also has a less comfortable side. A university should bring students into contact with people who disagree, require them to defend claims, and give them responsibilities that affect others. This can happen in a seminar, residence hall, laboratory team, student organization, or service project. The repeated practice of listening, revising, deciding, and accepting consequences is the point.

Artificial intelligence can simulate disagreement and offer feedback. It cannot create mutual obligation. A classmate, teacher, client, or patient has interests of their own and will still be present tomorrow.

Some learning requires contact with the world

A fluent explanation of a laboratory procedure is not a laboratory. A simulated patient is not always an adequate substitute for a frightened person. A generated design does not teach how materials fail, how a client changes the brief, or how a team recovers when a plan meets reality.

Simulation matters. It lowers cost, permits repetition, exposes students to rare cases, and allows failure where real failure would be dangerous. Its best use is often preparation for experience, not replacement of the experience that contains the educational point.

Laboratories, studios, clinics, archives, fieldwork, and collaborative projects force students to encounter resistance. The instrument drifts. The source is incomplete. The patient’s priorities differ from the textbook. The teammate misunderstands. Those are often the moments when judgment begins.

Universities do not own direct experience. Employers, apprenticeships, community organizations, museums, and independent projects may provide it more authentically. The university’s advantage is the ability to connect experience with disciplinary knowledge, supervision, reflection, and increasing responsibility.

Work matters, but it is not everything

The post-college world has changed. Graduates will enter organizations where artificial intelligence participates in research, analysis, communication, design, software development, customer service, and management. Colleges that ignore this environment leave students to discover new working conditions on their own.

Work preparation is therefore a real obligation. Students should learn to use powerful tools, verify their output, explain their decisions, collaborate across expertise, and remain responsible when a tool contributes to the result.

Employment cannot be the only purpose. A university also prepares people to interpret evidence, participate in civic life, make ethical choices, understand histories they did not live, and keep learning when a current job or tool disappears. A curriculum tied only to today’s job descriptions may become obsolete faster than one built around durable knowledge and adaptable judgment.

A credential is useful, but it is not the education

A degree helps employers make decisions without investigating every course, project, and applicant. Tyler Cowen describes the credential as a low-cost screening device when direct assessment is expensive.[2] That does not make the signal fraudulent. It makes it compressed.

Compression saves time and hides variation. Two graduates with the same degree may possess different capabilities, while two people with similar capabilities may not receive the same opportunity if only one has the recognized signal.

Artificial intelligence may make work samples, simulations, structured interviews, portfolios, and direct skill tests cheaper. Those alternatives can still be coached, gamed, biased, inaccessible, or expensive. The goal is not to replace one proxy with several fashionable ones. It is to reduce the distance between the credential and the capability it claims to represent.

Apply the substitution test

A student may combine community college, online study, employer training, independent projects, apprenticeship, professional certification, and artificial-intelligence tutoring. Another may need the structure, relationships, facilities, and protected time of a residential university.

The substitution test is function by function. Can an alternative develop durable knowledge, provide timely feedback and mentoring, create direct experience and increasing responsibility, support completion, and produce a credible signal without rebuilding the same cost and gatekeeping?

Price also requires care. Published tuition, cost of attendance, student net price, and institutional net tuition revenue are different numbers.[3][4] College Scorecard data add information about debt, repayment, and earnings, but observed earnings do not isolate what the institution caused and do not capture the whole value of education.[5]

Affordability, completion, debt, field, learning, relationships, and opportunity cost belong in the same decision. No single number settles the value of college, but complexity is not permission to hide the price.

The burden of proof runs both ways

Universities often describe their most expensive and least measurable practices as essential. Community, tradition, exploration, and mentorship can become noble language wrapped around weak teaching, administrative growth, or benefits available mainly to students who already know how to find them.

That objection is fair. A university should show where each promised benefit appears in the student’s experience, who receives it, what it costs, and what students can do because of it.

The same test applies to disruption. An online platform, employer academy, or artificial-intelligence tutor is not an adequate replacement just because it is scalable or inexpensive. Low cost is an advantage only if the result is worth having.

A value test for universities

Every institution should be able to answer four questions:

• What knowledge, capability, relationship, experience, and judgment does the institution promise, and where is each delivered?

• What evidence shows that students receive and can use those benefits?

• What does the benefit cost in money, time, debt, and foregone alternatives?

• What result would cause the institution to redesign, replace, or stop the practice?

The bargain universities must earn

I believe the university remains worth defending where it creates a disciplined community of learning that develops knowledge, judgment, relationships, experience, and responsibility over time. Age, selectivity, prestige, and authority to issue a credential are not enough.

Human attention and direct experience are costly. Good mentoring does not scale like a video or chatbot. If these are the university’s distinctive contributions, they should receive more money, time, and management attention.

The task is to strengthen the parts of the university that produce learning and formation while changing the parts built around information scarcity, opaque signals, and inherited routines.

Part three, The Goal Is Better-Chosen Knowledge, asks what must still live inside the learner. When explanations are always available, what knowledge must a person command?

Questions, corrections, or disagreements are welcome. You can reach me directly at dave@aworkingmodel.com.

Sources

1. Strada-Gallup, “2018 Alumni Survey: Mentoring College Students to Success.” Published 2018. National retrospective alumni survey; useful for reported mentoring prevalence and association, not a causal estimate of mentoring effects.

2. Tyler Cowen, “Against credentialism.” Marginal Revolution, Mar 18, 2022. Argument about education, signaling, and the cost of direct evaluation.

3. National Center for Education Statistics, “Price of Attending an Undergraduate Institution.” Updated May 2024. Official definitions for cost of attendance, grants, and net price.

4. National Association of College and University Business Officers, “Most Undergraduates Receive Grants at Private Colleges, NACUBO Research Shows.” Jun 1, 2026. Preliminary estimates from participating private nonprofit colleges; discount rate is not student net price or educational value.

5. U.S. Department of Education, “College Scorecard Field of Study Data Documentation.” Jun 2024. Methodology for program-level earnings, debt, and repayment measures; coverage and cohorts vary, and observed earnings are not a causal estimate of institutional contribution.

Generative artificial intelligence makes competent-looking intellectual work cheap. A student can produce an explanation, program, summary, or essay in seconds. The result may be insightful, mediocre, subtly wrong, or fabricated, and may also look better than the student’s understanding.

That change exposes an old weakness: a large gap between performance and learning. Much of education was built to move large numbers of students through lectures, assignments, examinations, grades, and credentials at manageable cost. Assessing performance in a deep learning environment is harder, more labor-intensive, and more expensive, requiring attention, practice, feedback, revision, direct experience, and judgment.

Artificial intelligence makes the gap impossible to ignore and provides an opportunity to address it.

The short answer

A polished answer is no longer reliable evidence that a student can produce, explain, or defend it.

The early research does not establish that artificial intelligence is generally good or bad for learning. It shows that the tool, the task, the timing of assistance, and the definition of learning can change the result.

Colleges should adapt quickly because the world graduates are entering has already changed. Their strongest response is not a universal policy but a clearer account of what students must learn and better evidence that they learned it.

Education was built around scalable proxies

A university education promises more than information. It promises knowledge, judgment, technique, sustained attention, and the ability to act when the answer is not obvious. Yet the current system rewards something narrower: produce the paper, pass the exam, earn the grade, collect the degree.

Each of those steps compresses a lot of information. A grade turns many judgments into one symbol. A transcript turns years of work into a pattern of symbols. A degree gives employers and other institutions a convenient signal when direct evaluation is costly.

The problem is not that proxies like grades, transcripts, and degrees exist; large institutions need ways to make decisions. The problem is that these proxies have become the objective. Students are incentivized to protect the grade, satisfy the rubric, choose safer work, and reproduce the expected form.

In short, be careful what you measure; you will get that.

The much less expensive system favors what can be delivered and scored at scale. The more demanding system asks students to discuss, build, investigate, revise, defend, and apply. Good teachers have often used laboratories, studios, fieldwork, oral examinations, mentoring, and repeated practice because those methods reveal how students think and what they can do independently.

Even when its use is permitted and disclosed, artificial intelligence makes the finished product even less informative, revealing little about what the student noticed, misunderstood, corrected, retained, or can now do alone.

Two experiments show why design matters

In a randomized high-school mathematics experiment, nearly 1,000 students at one school used ordinary classroom resources, a general GPT-4 interface, or a tutor designed with teacher-supplied solutions, common errors, and instructions to offer hints rather than answers.[1]

Both artificial-intelligence groups performed better during practice. After the tool was removed, students who had used the general interface performed below the control group. The designed-tutor group performed about the same as the control group. The general interface improved assisted work but weakened independent performance after the assistance disappeared.

A randomized crossover experiment in a Harvard physics course produced a different result. Students using a carefully engineered AI tutor earned higher immediate post-test scores than students in an active-learning class and reported greater engagement.[2] The tutor used expert-written prompts, solutions, videos, sequencing, and self-pacing.

I’m convinced that artificial intelligence can and will improve learning. Results will depend on how the tool is designed, when it intervenes, what expertise shapes it, and what students must still do themselves. The work will probably need to be adapted at several levels, from the institution and discipline to the course, assignment, and individual student. That is harder than distributing a generic chatbot, but it can produce much better learning at scale and at reasonable cost.

Define learning before choosing the tool

Any serious evaluation of learning should examine at least five outcomes, both with and without artificial-intelligence assistance:

• Performance while assistance is available.

• Independent performance after assistance is removed.

• Retention after time has passed.

• Transfer, meaning the ability to use what was learned in a materially different problem or setting.

• Judgment about when the tool is wrong, incomplete, or inappropriate.

A tool may improve one outcome and leave the others unchanged. It may save time, provide feedback, or help a student attempt harder work. It may also complete the intellectual step the student needed to practice.

This is why asking whether artificial intelligence is good for learning is too vague. An answer-giving chatbot, a tutor that withholds the answer, a simulation, and a critic that challenges an argument are different interventions. An immediate quiz, a delayed examination, an oral defense, and a new real-world problem are different measures.

We need to start by naming what students should know and be able to do. Then decide where assistance helps, where independent performance matters, and what evidence can distinguish the two.

Practical risks

Authorship rules built around the final artifact are becoming harder to enforce. Colleges need to provide assignments that make formulation, reasoning, revision, explanation, and responsibility visible.

Student interactions with artificial-intelligence tools may include education records, personal information, unpublished work, or material belonging to someone else. Federal guidance requires institutions to examine control, authorized use, redisclosure, and security when online services handle education records.[3] UNESCO frames these concerns within a broader call for human-centered educational design.[4]

Generative systems can also produce confident errors. The National Institute of Standards and Technology organizes artificial-intelligence risk around governing, mapping, measuring, and managing the system across its life cycle.[5] In education, that means someone must be able to check the output and must own the decision to use it.

A student who never develops the underlying knowledge cannot be the reliable human reviewer the system assumes. Engineering disciplines, including software development, often use separate quality-assurance reviews for the same reason: important work should be checked by people with the knowledge and responsibility to find errors.

Adapt because the outside world has already changed

Students are using these rapidly changing tools, and employers are reorganizing work around them.

Graduates will be expected to frame problems, organize context, judge outputs, detect failure, work across disciplines, and accept responsibility for decisions made with artificial-intelligence assistance. Colleges that fail to expose students to real-world environments will leave students unprepared.

Colleges can act now and improve their approaches over time. A college can change a course, name what students should learn, compare assisted and independent work, evaluate retention and transfer, protect student data, and revise the design when it fails.

A better diagnostic for colleges

The first question is not, “Do we allow AI?” It is, “What are students supposed to learn here?”

Then ask:

• What must the student understand and do without assistance?

• Where can artificial intelligence provide explanation, feedback, simulation, or productive challenge?

• How will the course evaluate retention, transfer, and judgment?

• What information enters the system, and who is responsible when something goes wrong?

• What result would cause the course to revise or stop the design?

Making these changes will require hard work and change management. The answers will differ across a literature seminar, nursing clinical, engineering laboratory, and accounting course. That is a feature, not a failure. The common requirement is a clear learning objective and honest evaluation within each course.

The work ahead

A finished answer is becoming cheaper. Deep learning is not. It still requires time, practice, feedback, experience, and human attention.

Performance is what a student can produce under stated conditions. Learning is the durable change in what the student knows, can use in a new setting, and can judge. Colleges and employers need evidence of both.

We do not yet know which designs will work best across subjects, institutions, and students. But colleges are preparing students for a world in which artificial intelligence is already part of serious work. They need to adapt now, evaluate the results, and keep improving the design.

Part two, What the University Is Still For, turns to the institution itself. If information and competent assistance are increasingly abundant, what is the university still for?

Questions, corrections, or disagreements are welcome. You can reach me directly at dave@aworkingmodel.com.

Sources

1. Bastani et al., “Generative AI without guardrails can harm learning: Evidence from high school mathematics.” Proceedings of the National Academy of Sciences, Jun 25, 2025; corrected Aug 20, 2025. Randomized field study at one Turkish high school over four short sessions.

2. Kestin et al., “AI tutoring outperforms in-class active learning.” Scientific Reports, Jun 3, 2025. Randomized crossover study of 194 students, two lessons, and immediate post-tests in one Harvard physics course.

3. U.S. Department of Education, Student Privacy Policy Office, FERPA FAQ on using online tools in a course. Current guidance accessed Sep 28, 2026. Operational privacy guidance, not a finding about learning.

4. UNESCO, “Guidance for Generative AI in Education and Research.” Published Sep 7, 2023; updated Jan 16, 2026. International guidance on privacy, human oversight, inclusion, and pedagogical design.

5. National Institute of Standards and Technology, “AI Risk Management Framework.” AI RMF 1.0 released Jan 26, 2023; Generative AI Profile released Jul 26, 2024. Voluntary risk-management guidance, not an educational-effectiveness study.

As a programmer at the University of South Carolina Computer Services Division, where I started in 1981, I gave visitors tours of the computer center, which supported the statewide public university education system. I remember the delivery and installation of a refrigerator-sized 1+GB IBM 3380 disk drive among the mainframe equipment; you can buy 1GB thumb drives at a grocery store these days. Years later, after starting my own business, I ran hosting infrastructure in colocation facilities; the machines had changed; the raised floors, cooling, racks, noise, electrical service, and the need for someone to maintain everything had not.

I watched data communications move from ten characters per second teletypes to 300- and 1,200-baud modems, T1 lines, and broadband. Storage and processing improved just as dramatically. What changed was not only what these systems could do, but what it cost to do it. Better price/performance made services economical that would once have been impractical.

As Internet use and cloud computing expanded, Google, Amazon, and others built the facilities to support them. Bitcoin mining brought another wave of demand for computing, electricity, and specialized chips. Miners pushed from ordinary processors toward purpose-built chips that did more work per watt. By 2019, estimates of Bitcoin’s annual electricity use were being compared with that of entire countries. Now AI is driving another expansion of the infrastructure.[26]

Better technology has regularly eased one bottleneck while enabling new uses that expose the next.

Computing infrastructure has never been weightless. Today’s AI capacity is another stage in its development, not a wholly new infrastructure. Some projects are larger and more concentrated than what came before. But the national debate about their electricity and water use too often ignores history, comparable uses, and local choices. I want to cut through that rhetoric and ask what bargain each host community can actually make.

What changed inside the building

By 2023, colocation and hyperscale facilities housed about 74 percent of U.S. servers, up from roughly 20 percent in 2014.[9] That shift was well underway before generative AI became a mass service. AI is changing the work performed in those facilities and, at some sites, the demands placed on their power and cooling systems.

A former bitcoin-mining site may already have land and substantial electrical connections, but its mining chips cannot simply run AI models. Training builds or updates an AI model; inference is what happens when people use that model to answer a question, generate an image, or carry out a task. Each use requires computation; a longer task may call on the model many times as it plans, acts, and checks its work. Training draws attention because it is intensive, but inference creates continuing demand as a service gains users and uses. I also expect agents that work through longer tasks and run for more hours to add to that demand. OpenAI reported in September 2026 that ChatGPT had 1.2 billion weekly users. That helps explain why operators are building capacity, and why a host community needs to understand the electricity and water required to serve it.[20]

Why electricity matters

Data centers need electricity to run computers and remove the heat they produce. U.S. data-center electricity use rose nearly 90 percent from 2000 to 2005 as Internet services, business computing, and server installations expanded. It rose 24 percent over the next five years and about 4 percent from 2010 to 2014. Digital activity did not stop growing. Virtualization and more efficient facilities delivered more service with less additional electricity. That history matters because innovation is now addressing AI bottlenecks in chips, models, software, cooling, and how we use existing capacity.[9][27]

How much electricity is that?

What is the difference between power and energy? Power is the rate at which electricity is used at a given moment, measured in megawatts or gigawatts. Energy is the total used over time, measured in megawatt-hours or terawatt-hours. A project’s requested power describes the grid connection it seeks; its annual energy usage depends on what is built and how steadily it runs.

Why do local megawatts matter? A facility drawing 100 megawatts continuously is a large single-site load, roughly equivalent to one-tenth of a large nuclear reactor running at full output. In a year, it would use roughly as much electricity as 84,000 average U.S. residential customers did in 2024. Concentrating that load at one connection can require new substations and lines. For the host community, the central questions are who builds them, who pays, and what value the project delivers in return.[11]

How large is U.S. data-center use today? Berkeley Lab estimated 192 terawatt-hours in 2024, about 4.7 percent of U.S. electricity. For scale, that is about one-eighth of residential electricity sales and roughly eleven times estimated electric-vehicle electricity use that year.[2][11][15]

How fast might it grow? Berkeley Lab estimates that U.S. data centers used 192 terawatt-hours in 2024; its 2030 reference forecast is 649, with scenarios from 521 to 843. The reference case implies about 22.5 percent annual growth from 2024 to 2030, compared with about 13.7 percent annually during the 90 percent rise from 2000 to 2005. The International Energy Agency reports that energy use per AI task fell by at least tenfold each year in recent years as hardware and software improved, even as more demanding uses such as reasoning, video generation, and agents became more common. Cheaper inference uses less electricity per task, but it also makes new uses economical. Efficiency, price, capability, and demand are all changing at once. I expect efficiency to bend the electricity-growth curve again, as it did after the early Internet expansion, even if total use continues rising for some time.[2][10]

Demand can also be flexible. Some computing work can move to hours when the grid is less heavily used, easing costly peaks. Brattle modeled more than $100 billion in U.S. consumer savings over a decade from better use of the existing grid; flexible demand is one way to get there. A community can ask whether a proposed facility will actually shift its load, and on what terms.[17]

Could today’s shortage become tomorrow’s surplus?

Companies build data centers, generation, and grid connections based on forecasts made years in advance. Meanwhile, inference may become much cheaper per task, and cheaper AI may create uses no one forecast. Capacity could briefly outrun demand, or new demand could absorb it. We have seen both forces in earlier technology buildouts. The question is how they will interact this time.

U.S. utilities built for peak demand in the 1970s, which turned out to be lower than forecast. A natural-gas plant construction boom in the early 2000s again left excess generating capacity in many places.[5] Internet backbone providers also laid enough fiber that a 2001 federal review found ample backbone capacity, even as connections closer to customers remained constrained.[22]

ERCOT’s planning case includes 25 gigawatts of additional data-center load by 2035, roughly the power of twenty-five large nuclear reactors operating at full output.[6][21] There is a practical dilemma: build too far ahead and some generation or grid capacity may sit idle; build too slowly and infrastructure may hold back useful services. The answer will vary by region, timing, and the terms under which utilities and customers commit to investment.

What about supply? I see the growth of renewables as an advantage: it diversifies electricity supply, and solar and wind can be paired with on-site generation and batteries where the site and economics permit. The electric-vehicle supply chain is helping scale batteries for stationary storage, which can ease short peaks and make variable generation more useful. The International Energy Agency projects that renewables will supply nearly half of the additional electricity data centers use globally between 2024 and 2030, while natural gas and coal together supply more than 40 percent. This buildout is likely to use several energy sources, not a single national recipe.[12]

Galaxy’s Helios campus in Texas shows what this evolution looks like on the ground. Originally developed for bitcoin mining, the site now has its first AI phase for CoreWeave. Galaxy reports roughly 200 megawatts of gross electrical capacity for that phase, including 133 megawatts for computing equipment; the rest covers cooling, power delivery, and other supporting systems. Its land, electrical access, and experience running a power-intensive site combine to meet AI demand.[1]

Where does the heat, and the water, go?

Almost all the electricity a computer uses ends up as heat. Cooling moves that heat from the chips to the outdoors.

How is the heat removed? Older rooms often used chilled air. Dense servers may use rear-door exchangers, liquid cold plates on chips, or immersion in a nonconductive fluid. In one immersion design, the fluid stays liquid; in another, it vaporizes and condenses. The final stage may use dry air, evaporative cooling, or a hybrid. A liquid or “closed-loop” system can therefore have low on-site water use, or it can ultimately rely on an evaporative tower.[3]

What should a community ask about water? How much would the whole campus withdraw and consume, particularly in a dry season? How much wastewater would it return, and what would it contain? Where does the source water come from, and what happens during drought? A dry-cooling choice may reduce site water consumption while raising electricity demand. Water used to generate that electricity may be consumed in a different watershed.

Water usage effectiveness, or WUE, measures on-site water use per kilowatt-hour used by computing equipment. Berkeley Lab modeled a U.S. average of about a tenth of a gallon per equipment kilowatt-hour in 2023. That average hides a wide range: heat rejected through dry air uses very little on-site water, while evaporative systems use water to reduce electricity needs. A hybrid can switch between the two as conditions change.[3]

Utah’s proposed Stratos campus combines computing, on-site power generation, and a water system in one large project. It shows why a community should negotiate the whole system, not just ask whether a server room uses water: the power plant, cooling equipment, treatment, and discharge can all shape the local bargain. Box Elder County and Utah have tools to decide whether the proposed terms work for them. Local jurisdictions' ability to set terms is a strength of the American approach, not an obstacle to remove.[4]

Could cheaper electricity make water cheaper? Yes, by lowering pumping and desalination costs. Then treatment, pipelines, and concentrated brine become the next bottlenecks. Once one constraint eases, solving the next often becomes more attractive. Stratos is a useful test of whether that process can work inland, where water quality and concentrate disposal matter.[7]

What are we choosing to scrutinize?

How much freshwater do familiar activities consume? Counting the electricity that runs them, Berkeley Lab estimates U.S. data centers used about 230 billion gallons in 2023. A rough estimate for U.S. golf courses in 2024 is about 400 billion gallons. A beef-production model estimates about 6.1 trillion gallons per year, mostly for irrigating feed crops. That puts golf in the same broad range as data centers on this rough estimate, while beef production is far larger.[13]

Why do we scrutinize one use so intensely while treating other familiar uses as ordinary? Why do some loud voices claim that water use from computing infrastructure growth is uniquely objectionable? We need a consistent standard: what water is being used, what value that use produces, and what alternatives the community has. When we single out a new use while overlooking familiar ones, our prior beliefs may do more of the work than the numbers.[16]

The next constraint

In The AI Industrial System, I described how pressure can shift from one part of the AI system to another. For a data-center project, the grid connection may be the immediate constraint; elsewhere, water, heat removal, equipment, labor, or community consent may come next. Those pressures are already driving responses: on-site power generation, immersion and on-chip liquid cooling, and new ways to manufacture gas-turbine blades. I expect some attempts to fail and others to work. But pressure to solve these engineering constraints comes from more than today’s AI services. Workforce changes may strengthen demand for the computing they support.[3][24]

Many societies are also facing a shrinking pool of working-age people. The United Nations says population has already peaked in 63 countries and areas, including China, Germany, and Japan, even though the world as a whole is still growing. Across the OECD, the population aged 20 to 64 is projected to fall 8 percent between 2023 and 2060; in a quarter of member countries, the projected decline exceeds 30 percent. China’s own statistics recorded 3.39 million more deaths than births in 2025, continuing its population decline. Those trends do not mean every occupation faces a shortage or that software can simply replace people. They do mean that aging societies have reason to make their workers more productive. I expect that pressure to increase investment in robotics and AI, making computing demand a response to a real economic constraint, not just a technology fad.[23][25]

The measure that matters is useful work delivered for the energy, water, and capital invested, with humans still accountable for how the systems are used.[14][23]

The local choice

Data centers cluster where power, fiber, land, customers, and favorable terms come together, a pattern evident in Northern Virginia, the world’s largest concentration of data centers. Existing networks and expertise attract additional facilities, and that concentration can increase pressure on shared utilities. Texas, Utah, and Georgia offer different combinations of these ingredients. Jurisdictions cannot be forced to accept data center construction they don’t want, and that’s as it should be.[18]

The jurisdictional and regulatory framework for data center construction is not new: land-use rules, water rights, utilities, environmental protections, and contracts already shape what can be built. Municipalities can accept, condition, or reject projects where they hold that authority; counties and state bodies decide some projects elsewhere. Variation among jurisdictions creates a marketplace, giving developers choices and communities leverage. This is the American way!

Atlanta shows the local bargain in action. The city restricted new data centers in certain transit-oriented areas, and Digital Realty is seeking an exception for a West End site. Whatever happens to that request, the case shows a city making its own judgment about where a data center belongs and what terms it should meet.[8]

Some projects will be good bargains and some will not. A project that shifts grid-upgrade costs onto households or strains scarce dry-season water could be a bad bargain. In investment banking, I learned to examine the terms before accepting the headline. Communities can ask what infrastructure must be built, who pays, what water is needed, what benefits are plausible, and which obligations are enforceable. Tax incentives may help win a project, but they are a price the jurisdiction chooses to pay, not a substitute for permission to build.[19]

One community may welcome a data center for its tax base and infrastructure; another may reject a particular parcel or water source. Existing state and federal protections still apply where required. If residents dislike their rules, they can change them through the responsible jurisdiction’s own process. That is how self-government should work.

I expect those decisions to shape the data center rollout, not stop it. Demand for computing has driven successive waves of infrastructure, and AI is just the next wave. Places that plan well, including economically struggling communities seeking investment, can compete for projects and ask for favorable terms. The long-term value will show up in tax receipts, utility costs, water use, jobs, and complementary activity, not in the scale of a press release.

The cloud has always lived somewhere. I saw one of its earlier homes through glass at Georgia Tech. Today’s campuses are another stage of that evolution. Communities can decide what to welcome, what to require, and what to refuse. The stronger the demand, the more important it is to be clear about the bargain.

You can reach me directly at dave@aworkingmodel.com.

Sources

[1] Galaxy, Helios phase-I release, July 6, 2026: https://www.galaxy.com/newsroom/galaxy-completes-phase-i-of-its-helios-data-center-campus ; Galaxy, mining-site repurposing, May 14, 2025: https://www.galaxy.com/newsroom/building-the-infrastructure-for-an-ai-powered-future.

[2] Lawrence Berkeley National Laboratory, United States Data Center Energy Usage Report: 2025 Update, June 2026: https://www.energy.gov/documents/united-states-data-center-energy-usage-report-2025-update . Model and scenario, not metered AI-only energy.

[3] U.S. Department of Energy, Best Practices Guide for Energy-Efficient Data Center Design, 2024: https://www.energy.gov/sites/default/files/2024-07/best-practice-guide-data-center-design_0.pdf ; DOE, Cooling Water Efficiency Opportunities, 2019: https://www.energy.gov/cmei/femp/cooling-water-efficiency-opportunities-federal-data-centers ; Lawrence Berkeley National Laboratory, 2024 U.S. data-center report: https://escholarship.org/content/qt32d6m0d1/qt32d6m0d1.pdf . Site WUE is annual on-site water per IT kWh; Berkeley Lab’s modeled 2023 average is just over 0.36 L/IT-kWh. Its 4.52 L/kWh indirect estimate uses total facility electricity and regional grid-mix assumptions; it is not site withdrawal, and the two denominators must not be added as written. Cooling designs trade site water against electricity and differ by climate and operation.

[4] Box Elder County, Stratos FAQ, accessed September 25, 2026: https://www.boxeldercountyut.gov/661/Stratos-Page-FAQ-Page ; State-hosted FAQ: https://governor.utah.gov/wp-content/uploads/FAQ-on-Stratos-Project.pdf ; MIDA Resolution 2026-06, April 24, 2026: https://www.midaut.org/s/20260424-Stratos-Project-Area-Plan-MIDA-BOARD-Resolution-06-dggg.pdf ; county–MIDA agreement, May 4, 2026, section 5(j): https://www.boxeldercountyut.gov/DocumentCenter/View/2189/Box-Elder-County-MIDA-Stratos-Agreement-26-15-.

[5] U.S. Energy Information Administration historical capacity assessment: https://www.eia.gov/electricity/pdfpages/perfiss/index.php ; early-2000s gas buildout: https://www.eia.gov/todayinEnergy/detail.php?id=1690 ; California curtailment: https://www.eia.gov/todayinEnergy/detail.php?id=65364 . Distinct kinds of surplus.

[6] North American Electric Reliability Corporation, 2025 Long-Term Reliability Assessment, Texas RE-ERCOT section, published 2026: https://www.nerc.com/globalassets/our-work/assessments/nerc_ltra_2025.pdf . Planning forecast and its adjustments; no claim of a realized 2035 outcome.

[7] Bureau of Reclamation, Estimating the Cost of Brackish Groundwater Desalination in Texas, July 2014: https://www.usbr.gov/gp/otao/estimating_cost_brackish_groundwater_desalination_texas.pdf ; Texas Water Development Board desalination FAQ: https://www.twdb.texas.gov/innovativewater/desal/faq.asp . Planning-level cost drivers from Texas, not a Stratos estimate.

[8] Atlanta City Council, 2024 data-center zoning action: https://citycouncil.atlantaga.gov/Home/Components/News/News/3928/175?arch=1 ; Digital Realty, Atlanta West End project page, accessed September 26, 2026: https://www.digitalrealty.com/about/projects/atlanta-west-end ; NAACP Legal Defense Fund, August 21, 2026 opposition letter: https://www.naacpldf.org/wp-content/uploads/Atlanta-City-Council-data-center-letter-8.17-FINAL-2.pdf.

[9] Lawrence Berkeley National Laboratory, United States Data Center Energy Usage Report, 2024: https://escholarship.org/content/qt32d6m0d1/qt32d6m0d1.pdf ; Lawrence Berkeley National Laboratory, United States Data Center Energy Usage Report: 2025 Update, June 2026: https://www.energy.gov/documents/united-states-data-center-energy-usage-report-2025-update ; Lawrence Berkeley National Laboratory, 2016 assessment: https://bies.lbl.gov/publications/united-states-data-center-energy . Modeled electricity and server-location shares, not a consistent historical series of facility floor area; report vintages revise prior estimates.

[10] International Energy Agency, Key Questions on Energy and AI, April 2026: https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary . Global growth rates are IEA estimates, not U.S. project-meter readings; AI-focused and all-data-center categories differ.

[11] U.S. Energy Information Administration, estimated light-duty EV electricity use, 2024, table D1: https://www.eia.gov/electricity/monthly/current_month/april2026.pdf ; U.S. Energy Information Administration, average U.S. residential electricity purchases in 2024: https://www.eia.gov/todayinenergy/detail.php?id=65244 . EV and household figures are national annual energy comparisons; neither gives local coincident peak demand. The 100 MW example assumes continuous full load for 8,760 hours.

[12] International Energy Agency, Energy and AI, April 2025: https://www.iea.org/reports/energy-and-ai/energy-supply-for-ai ; IEA, Electricity 2026, flexibility section: https://www.iea.org/reports/electricity-2026/flexibility ; IEA, Global EV Outlook 2026, EV batteries section, May 2026: https://www.iea.org/reports/global-ev-outlook-2026/electric-vehicle-batteries . Supply shares and battery deployment are projections or market observations, not proof of any P007 site’s generation mix or storage duration. EV battery terawatt-hours measure storage capacity, not annual electricity consumption.

[13] Lawrence Berkeley National Laboratory, 2024 data-center report: https://escholarship.org/content/qt32d6m0d1/qt32d6m0d1.pdf ; Golf Course Superintendents Association of America, 2025 water survey: https://www.gcsaa.org/docs/default-source/what-we-do/gcep-phase-4-water-report.pdf?sfvrsn=c829dd3e_0 ; Rotz et al., U.S. beef farm-gate water-footprint model, 2019: https://orcattle.com/wp-content/uploads/2021/02/Footprint-of-US-Beef-Production.pdf . For a consumption-based comparison, Berkeley Lab estimates that in 2023, data centers used 66 billion liters directly and nearly 800 billion liters indirectly for electricity, or about 230 billion gallons combined. The GCSAA estimates 1.63 million acre-feet of golf irrigation in 2024; applying a historical 75% golf-course consumptive-use coefficient cited by the U.S. Geological Survey (https://pubs.usgs.gov/sir/2007/5197/) gives roughly 400 billion gallons. That coefficient comes from a Great Lakes study, so the national golf figure is approximate. Rotz et al. model 23.2 trillion liters of annual blue-water consumption for U.S. beef production, or about 6.1 trillion gallons, mostly from feed-crop irrigation. The estimates span different years and activity boundaries.

[14] IEA, Key Questions on Energy and AI, April 2026: https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary ; U.S. Bureau of Labor Statistics, labor-force projection 2024–34, September 2025: https://www.bls.gov/careeroutlook/2025/data-on-display/economy-and-labor-force.htm . Energy per task varies enormously; the labor projection is context for possible demand, not proof of AI adoption or facility utilization.

[15] U.S. Energy Information Administration, Electric Power Annual, table 2.8, final 2024 retail sales: https://www.eia.gov/electricity/annual/table.php?t=epa_02_08.html . Residential: 1,482.874 TWh; industrial: 1,034.584 TWh. Data-center electricity is not an extra sector to add to those sales; EIA notes classification and reporting changes for data centers and mining in commercial and industrial sales.

[16] Ziva Kunda, “The Case for Motivated Reasoning,” Psychological Bulletin, 1990: https://pubmed.ncbi.nlm.nih.gov/2270237/ ; Raymond S. Nickerson, “Confirmation Bias: A Ubiquitous Phenomenon in Many Guises,” Review of General Psychology, 1998: https://doi.org/10.1037/1089-2680.2.2.175 ; Valerie F. Reyna and C. J. Brainerd, “Numeracy, Ratio Bias, and Denominator Neglect in Judgments of Risk and Probability,” Learning and Individual Differences, 2008: https://www.sciencedirect.com/science/article/pii/S1041608007000428 . These sources describe general tendencies, not the motives of any named P007 opponent or developer. A national denominator may supply perspective but cannot settle a local water or power constraint. The symmetrical questions are an editorial test, not a psychological diagnosis.

[17] Brattle, The Untapped Grid (2026). Illustrative utility-system modeling, not observed savings or proof of flexibility at any named site.

[18] Virginia Joint Legislative Audit and Review Commission, Data Centers in Virginia. Identifies reasons for Northern Virginia’s concentration, not the cause of every project elsewhere.

[19] Texas governor’s tax-incentive proposal; Texas Comptroller’s qualifying-data-center exemption guidance; Pennsylvania Executive Order 2026-05 announcement. The Texas phase-out is proposed; Pennsylvania directed conditions on an existing exemption.

[20] OpenAI, DevDay 2026 Recap (September 29, 2026): https://openai.com/index/devday-2026-recap/ . Company-reported 1.2 billion weekly users.

[21] U.S. Department of Energy, “How Much Power Does a Nuclear Reactor Produce?” (March 31, 2021): https://www.energy.gov/ne/articles/infographic-how-much-power-does-nuclear-reactor-produce . A typical reactor is approximately one gigawatt of generating capacity.

[22] U.S. Government Accountability Office, Telecommunications: Characteristics and Competitiveness of the Internet Backbone Market (October 16, 2001): https://www.gao.gov/products/gao-02-16 . The review found ample backbone fiber but identified possible local access constraints.

[23] OECD, Employment Outlook 2025 (July 9, 2025): https://www.oecd.org/en/about/news/press-releases/2025/07/oecd-job-markets-remain-resilient-but-population-ageing-will-cause-significant-labour-shortages-and-fiscal-pressures.html ; United Nations, World Population Prospects 2024 summary (July 11, 2024): https://www.un.org/sustainabledevelopment/blog/2024/07/press-release-wpp2024/ ; International Federation of Robotics, World Robotics 2026 release (September 24, 2026): https://ifr.org/ifr-press-releases/five-million-robots-now-operate-in-factories-globally . Population and robot-installation trends; the effect on AI demand is Dave’s forecast.

[24] Siemens, 3D-printed gas-turbine blade tests (February 6, 2017): https://press.siemens.com/global/en/pressrelease/siemens-achieves-breakthrough-3d-printed-gas-turbine-blades ; Jennifer Hiller, The Wall Street Journal, Is 3-D Printing the Answer to AI’s Gas Turbine Crunch? (September 24, 2026): https://www.wsj.com/business/energy-oil/is-3-d-printing-the-answer-to-ais-gas-turbine-crunch-865ae5d0 . A component test and a reported attempt to accelerate turbine-blade production.

[25] National Bureau of Statistics of China, Statistical Communiqué on the 2025 National Economic and Social Development (February 28, 2026): https://www.stats.gov.cn/english/PressRelease/202602/t20260228_1962661.html. Official 2025 population, births, deaths, and age distribution; the forecast that demographic pressure increases AI demand is Dave’s interpretation.

[26] Cambridge Centre for Alternative Finance, Cambridge Bitcoin Electricity Consumption Index launch, July 2, 2019: https://www.jbs.cam.ac.uk/2019/cambridge-bitcoin-electricity-consumption-index-cbeci/ ; Cambridge Judge Business School, Bitcoin electricity consumption: an improved assessment, 2023: https://www.jbs.cam.ac.uk/2023/bitcoin-electricity-consumption/ . The 2019 annualized estimate and country comparisons give historical scale; the later review describes specialized mining hardware and revised estimates.

[27] U.S. Environmental Protection Agency, Report to Congress on Server and Data Center Energy Efficiency, 2007: https://datacenters.lbl.gov/sites/default/files/EPA_Datacenter_Report_Congress_Final1.pdf . Contemporary account of Internet services, business digitization, record retention, and disaster recovery as contributors to data-center growth; it does not apportion the 2000–2005 electricity increase.

We wanted to make that easier.

Instead of forcing people to learn a computer system’s language and structure, we wanted the system to adapt to how people naturally communicate. We built a working architecture that combined speech recognition, natural language processing, semantic models, text-to-speech, wireless devices, and contextual information from technologies such as radio-frequency identification and GPS. A user could ask a question in ordinary English; the system would determine what the person meant, translate that request into a database operation, retrieve the relevant information, and return a direct answer, potentially by voice and without the user ever seeing a conventional computer interface.

One of our first demonstrations was a virtual museum docent. A visitor could stand near a work of art and ask questions about it. Because the system knew the visitor’s proximity to a particular artwork, the spoken question didn't need to include every piece of information needed to answer it. The software could combine the words with external context to determine which object the visitor meant. The original proof of concept was built in 2004 and publicly demonstrated the following year.

The underlying technology was very different from today’s transformer-based generative AI, and I do not regard it as an early version of ChatGPT. It was not a large language model, a generative AI system, an agent, or a retrieval-augmented generation system. What strikes me in retrospect is something more interesting: many of the practical problems we were trying to solve are the same ones developers face today when they try to turn an impressive AI model into a reliable, useful system.

The models changed dramatically. Many of the underlying problems did not.

Context matters as much as language

Human language is inherently ambiguous. Meaning depends not just on words, but on context, specialized terminology, prior interactions, location, and shared assumptions. Our approach relied heavily on semantic models that represented the particular business or application domain. We deliberately constrained the domain rather than trying to solve the entire problem of understanding human language, making the system more accurate, more predictable, and much more practical to deploy. We considered daisy-chaining several domain-specific systems into a larger, more powerful whole, but never got around to it.

Modern large language models start from an extraordinarily more capable foundation, but applied AI still depends heavily on supplying the right context. Today’s terminology includes grounding, retrieval, persistent context, domain knowledge, memory, and knowledge architectures. The mechanisms have changed, but the practical question is recognizable: What information does the system need in order to understand what this person means in this particular situation?

A second challenge was turning language into action. We distinguished what we called discrete answer retrieval from conventional document retrieval. If someone asks, “Who are our top ten customers?”, the desired result is not a collection of documents that may contain the answer. The system has to understand the request, map the language to the underlying data model, construct the appropriate database query, execute it, and return the answer. Our platform was explicitly designed to translate natural-language concepts into operations against structured information systems.

This parallels today’s AI systems that translate a user’s request into database queries, searches, application actions, or tool calls. I would not equate the technologies; current models can operate across language and information in ways that were simply unavailable to us. But the architectural problem, moving from what the user says to what the system should do, is essentially the same.

We also concluded early that language alone was sometimes insufficient. One idea that emerged from the project, and later became part of our patent portfolio, was to combine a natural-language request with information from other sources. Location and proximity were obvious examples. If someone standing in front of an object asks, “Who made this?”, the word this is meaningless unless the system can understand the surrounding context. We used RFID tags attached to paintings to establish proximity.

That led us toward a multimodal architecture. Voice, location, RFID, GPS, and other environmental inputs could all help interpret the request. The goal was not simply better speech recognition. It was better intent understanding by combining what a person said with what the system knew about the situation.

Today’s multimodal AI systems can work with language, images, audio, application state, external data, location, and other forms of context on a scale we could not have approached in 2004. Yet one central design question remains unchanged: Which pieces of context should the system consider, and how should they influence its interpretation of what the user wants?

We also had to decide what happened when the machine was wrong

No semantic model could anticipate every question. Rather than assuming the system would always understand, we designed around failure. If it understood most of a request but encountered an ambiguous portion, it could ask the user for clarification. If it could not resolve the problem itself, a human could step in. We also considered retaining successful interactions and regression-testing them after semantic-model changes so that fixing one problem did not silently break answers that had previously worked.

In another proposed application, the system would automatically resolve requests the natural-language engine could handle confidently, while routing harder cases to a conventional human customer-service workflow. The expectation was that the proportion handled automatically would increase as the technology, semantic model, and knowledge repository improved.

Today we use terms such as human-in-the-loop, escalation, confidence thresholds, evaluation, regression testing, and continuous improvement. The vocabulary is newer, and the systems are immeasurably more capable, but the operational lesson is familiar: useful AI is not simply about producing an answer. It also requires knowing when the answer may not be good enough, testing the system systematically, and deciding when people need to remain involved.

The work led to four U.S. patents

That work led to four issued U.S. patents in multimodal natural-language query technology, on which I was the named inventor: U.S. Patent Nos. 7,376,645; 7,873,654; 8,150,872; and 9,223,776.[1] The first patent claims priority to November 29, 2004, was filed on January 24, 2005, and was issued on May 20, 2008. The later grants extend the portfolio through continuation applications and later claims.

I value that work today not because it proves that we somehow foresaw large language models (we did not). I value it because it offers a useful perspective on how much has changed, and how much has not. The long-standing ambition has been to make computers understand what people actually want without forcing people to learn how computers represent the world. What has changed is our ability to get much closer to that goal.

From custom language systems to modern AI

In 2004, accomplishing even a relatively constrained version of this required considerable custom engineering. We had to worry about recognition grammars, semantic models, application-specific vocabularies, database schemas, speech technologies, integration, and specialized logic. Our planning materials called for continuing work on semantic models, statistical language techniques, scalability, administration tools, metrics, security, and exception handling.

Modern foundation models greatly reduce the custom language engineering required to get started. A capable general-purpose model can start with linguistic abilities that once would have required a major development effort. That changes the economics of the problem enormously. It also shifts much of the difficult applied work away from building the language capability itself and toward the architecture surrounding the model.

Technological progress often looks discontinuous from a distance. A new capability suddenly becomes visible, adoption accelerates, and it seems the world changed almost overnight. But what looks like a step change often reflects decades of iterative work that crosses several thresholds at roughly the same time. Algorithms improve. Compute becomes more capable. Specialized hardware makes certain calculations practical. Data becomes more abundant and better organized. Cloud infrastructure makes capacity easier to obtain. Software frameworks reduce the cost of experimentation. Interfaces make the capability legible to ordinary users. Business models give firms a reason to keep investing, deploying, and improving.

Generative AI is a striking example, but the sequence was neither automatic nor inevitable. Large labeled datasets such as ImageNet made it possible to compare systems at a scale that smaller datasets could not support.[2] AlexNet’s 2012 result combined a deep neural network, a large dataset, and efficient GPU implementation to deliver a measurable performance gain.[3] NVIDIA introduced CUDA in 2006, making GPU throughput more accessible to general-purpose software.[4] The 2017 transformer architecture relied on attention rather than recurrence and was designed to be more parallelizable.[5] GPT-3 then showed how performance and few-shot behavior could change with much greater model scale, while also documenting important limitations in training data and evaluation.[6]

The surrounding ecosystem changed too. Amazon EC2’s 2006 beta offered resizable computing capacity billed by use rather than requiring every developer to own the underlying servers.[7] TensorFlow’s 2015 open-source release and later frameworks lowered the cost of building, sharing, and deploying machine-learning systems.[8] In 2020, the OpenAI API offered a general text-in, text-out interface, allowing developers to incorporate model capability into products without building the base model themselves.[9] ChatGPT’s conversational research preview in November 2022 made a much more capable language interface directly accessible to a broad public and invited rapid feedback through ordinary use.[10] Those delivery models also gave developers and users practical ways to buy access without financing the entire technical stack themselves.

None of those events alone explains generative AI’s emergence at scale. Nor does the sequence prove that the outcome had to occur. Technical progress can stall; economics can fail; regulation, energy, supply chains, organizational readiness, trust, and user behavior can redirect what gets built. The better claim is cumulative and contingent: a visible step change can occur when technical capabilities and an economically viable delivery system become good enough together.

I have treated the physical, financial, and institutional side of that delivery system in The AI Industrial System.[11] The narrower point here is historical. The public breakthrough was real in its effects without being discontinuous in its origins. A model architecture mattered, but so did the semiconductor and cloud capacity that trained and served it; the datasets that shaped it; the software tools that let teams experiment; the interfaces that let people use it; and the business arrangements that funded the whole system. Isolating the exact contribution of each is difficult, and their interactions matter more than a single-cause story.

The questions I spend more time thinking about today are therefore not simply about model intelligence. They concern what context an AI should receive, which information sources it should trust, what tools it should be allowed to use, how we determine whether its answers are good enough, and when it should ask for clarification or escalate to a person. They also concern autonomy and evaluation: how much freedom should the system have, and how do we improve weak behavior without damaging what already works? For a business, all of this eventually reduces to another question: Does the resulting system create enough value to justify its cost, complexity, and risk?

Those are questions of applied AI and AI strategy, not simply model capability.

A principle that still holds

Looking back, one principle from that early work has held up particularly well: technology should adapt to the person rather than forcing the person to adapt to the technology. We described our objective then as letting subject-matter experts access sophisticated information in the language they already understood, without needing to know the database, query language, or software architecture underneath.

Today’s AI lets us take that idea far beyond anything we could build two decades ago. I am not nostalgic for the old technology; modern models are incomparably more capable. What interests me is the problem’s continuity. The models changed, the interfaces changed, and the economics changed, but people still know what they want to accomplish, and computers still need a mechanism to translate that human intention into something the machine can reliably do.

Applied AI increasingly makes that translation the machine’s problem rather than the user’s. That was what interested me in 2004, and it is a large part of what interests me again today. In that respect, 2026 looks surprisingly familiar from a 2004 perspective.

Source note

This essay draws on my experience leading the project, surviving technical and demonstration materials from 2004 through 2006, the resulting patent record, and primary sources on the later development of modern AI.

Questions, corrections, or disagreements are welcome. You can reach me directly at dave@aworkingmodel.com.

Sources

1. U.S. Patent Nos. 7,376,645, 7,873,654, 8,150,872, and 9,223,776, multimodal natural-language query patents, issued 2008, 2011, 2012, and 2015.

2. Jia Deng et al., “ImageNet: A Large-Scale Hierarchical Image Database,” CVPR 2009.

3. Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, “ImageNet Classification with Deep Convolutional Neural Networks,” NeurIPS 2012.

4. NVIDIA, CUDA Programming Guide, “Introduction.”

5. Ashish Vaswani et al., “Attention Is All You Need,” 2017.

6. Tom B. Brown et al., “Language Models are Few-Shot Learners,” 2020.

7. Amazon Web Services, “Announcing Amazon Elastic Compute Cloud (Amazon EC2) - beta,” August 24, 2006.

8. Google, “TensorFlow: smarter machine learning, for everyone,” November 9, 2015.

9. OpenAI, “OpenAI API,” June 11, 2020.

10. OpenAI, “Introducing ChatGPT,” November 30, 2022.

11. Dave Bernard, “The AI Industrial System,” A Working Model, September 2026.

TL;DR

Artificial intelligence (AI) may arrive through a screen, but an industrial system builds it. Capital and electricity move through semiconductor equipment, fabrication, packaging, memory, networking, data centers, cloud platforms, models, applications, and human organizations before becoming useful intelligence. That system has genuine points of fragility, but its interdependence also creates strong incentives to solve constraints. Competition can weaken the economics of particular companies while strengthening the system as a whole. Falling costs may increase total demand rather than reduce it. The system can therefore become economically important even if some companies, projects, and investors fail. The right unit of analysis is not the model, the chip, or the data center alone. It is the complete system, including the people who finance, build, regulate, deploy, and use it.

Software never stopped being physical

Over five decades in software, I watched abstraction hide more and more machinery. Developers moved from controlling individual bytes and devices to calling libraries, services, and application programming interfaces. Business users moved from specialized terminals to personal computers, browsers, mobile devices, and cloud applications. Each generation made an expanding technical system feel simpler.

That was the point. The best software systems hide complexity so that people can concentrate on the problem they are trying to solve.

AI takes this abstraction much further. A user can ask an ordinary-language question and receive an answer assembled by a system whose physical and organizational complexity is almost entirely invisible. The interaction may feel weightless, but the system behind it is not.

A graphics processing unit (GPU) order is not an AI service. The processor must be designed, fabricated, packaged, supplied with high-bandwidth memory (HBM), connected to other processors, installed in data-center servers, cooled, powered, financed, programmed, and incorporated into a model. That model must then be delivered through a cloud service or application, connected to useful information and workflows, governed by an organization, and used by a person with a purpose.

AI is a physical industrial system disguised as software.

This distinction changes how we evaluate both opportunity and risk. A surprising model demonstration tells us little about how quickly reliable capacity can be manufactured, energized, deployed, and converted into economic value. Conversely, a delay at one data center does not mean the wider system has failed. The unit of analysis must be large enough to include the dependencies and specific enough to identify where the constraint has moved.

The system has five connected layers

The full system is complicated, but it becomes manageable when grouped into five layers.

1. Physical inputs include electricity, land, water or alternative cooling, raw materials, buildings, construction, and transmission capacity.

2. Compute infrastructure includes semiconductor equipment, fabrication, advanced packaging, processors, memory, networking, servers, storage, and data centers.

3. Platforms include cloud services, foundation models, development tools, security, evaluation systems, and model-routing software.

4. Applications and organizations turn technical capability into completed work through workflows, proprietary information, domain expertise, controls, and human judgment.

5. Capital and institutions surround every layer through financing, insurance, regulation, permits, standards, education, and public acceptance.

The layers are not a simple assembly line. They are a feedback system. Better models increase demand for computing. Greater demand attracts capital. Capital funds manufacturing and infrastructure. Additional capacity lowers cost and expands access. Broader use reveals new applications and new constraints, beginning another cycle.

This resembles Amazon’s long development arc. Retail growth that originally focused on a few books grew to require warehouses, logistics, software, data, operating disciplines, and large amounts of capital. Capabilities built for Amazon’s own use later became infrastructure for other businesses, most visibly through Amazon Web Services (AWS). AI is following a similar pattern at a much larger physical scale. Internal tools, data pipelines, evaluation methods, security controls, and trained people can become reusable infrastructure for the next application.

The earlier paper Compounding Intelligence describes this effect within a person or organization. This paper describes the industrial system that supplies the capacity on which that compounding depends.

Bottlenecks move

Every major system has constraints. The mistake is to assume that today’s constraint will remain the decisive one.

At one point, leading-edge processors appeared to be the central scarcity. More processors increased demand for high-bandwidth memory, advanced packaging, network switches, optical links, servers, cooling equipment, transformers, turbines, electricians, and grid connections. Solving one bottleneck raised the value of the next scarce input.

Some constraints are harder to replace than others. I use the word lynchpin for a component, capability, or organization whose prolonged loss would stop or materially delay the system because no substitute could match the required quality, scale, cost, and speed.

ASML Holding N.V. (ASML) is the clearest equipment example. Its extreme-ultraviolet lithography systems are essential to manufacturing the most advanced chips, and no competitor can presently supply an equivalent system at commercial scale. Existing machines would continue operating if new deliveries stopped, but maintenance, spare parts, upgrades, new capacity, and progress to later manufacturing generations would become increasingly constrained. ASML is therefore a technology-progress and future-capacity lynchpin, not an instant on-off switch. [3]

Taiwan Semiconductor Manufacturing Company (TSMC) is a manufacturing lynchpin. Its importance comes from leading-edge fabrication, yield, scale, advanced packaging, design support, and a broad customer ecosystem. Competing chip architectures can all depend on the same foundry. A major interruption would therefore affect many ostensibly independent companies at once. [4]

Other system components are deployment-rate constraints rather than system-essential lynchpins. Alternative suppliers or locations exist, but they cannot be added quickly or cheaply enough to preserve the original schedule. High-bandwidth memory, networking equipment, turbines, transformers, switchgear, cooling, skilled construction labor, and grid connections often fall into this category.

This distinction matters. A lynchpin can threaten the system’s technical path. A deployment constraint usually changes timing, cost, geography, and who captures the profit. Treating every shortage as an existential threat obscures the more practical question: what will the system do next?

Interdependence can create commitment

Complexity is often described only as fragility. That is half the story.

My work in investment banking and transaction due diligence trained me to look beyond a project’s promoter. I want to know who the counterparties are, what they have committed, how much they can lose, whether contracts are enforceable, and whether the participants have the technical and financial capacity to solve a problem when the plan changes. Promotional claims are cheap, but signed obligations by capable parties are not.

The AI buildout involves some of the world’s largest technology companies, semiconductor manufacturers, infrastructure managers, banks, utilities, contractors, sovereign investors, and institutional customers. Their participation does not prove that demand forecasts are correct or that every project will earn an acceptable return; capital has made spectacular mistakes before. It does show that the system is supported by parties with money, customers, technical staff, political influence, and strong incentives to remove obstacles.

I call this network commitment. It differs from common-mode dependency, in which many apparent participants ultimately rely on the same weak customer, financing source, technology, or assumption. A well-supported network can adapt when one route closes. A common-mode dependency can fail in several places at once because the apparent diversity was illusory.

This distinction matters especially in ecosystem financing. A supplier may invest in a customer, guarantee capacity, extend credit, or help assemble a project. Those arrangements can solve a legitimate coordination problem. They can also conceal demand that exists mainly because the supplier financed the purchase.

The useful questions are concrete. Is capacity being used? Are independent customers producing enough cash to pay for it? Does the contract last as long as the financed asset? Can the asset serve another customer? Who owns the downside if utilization disappoints? The label circular financing answers none of these on its own.

Demand should be measured in useful work

Tokens are useful for operating a model, but they are a weak measure of economic demand. A business does not ultimately want tokens. It wants a support case resolved, a deal room analyzed, working software produced, a molecule designed, a claim reviewed, a sales lead qualified, or a workflow completed.

The better unit is a useful task. Its measurement should include completion quality, time saved, human review required, error cost, latency, energy use, and economic value.

Current use is already large. OpenAI reported in September 2026 that its products reached more than one billion weekly active users and 2.5 million businesses. Anthropic reported in May that its run-rate revenue had crossed $47 billion. Microsoft Corporation (Microsoft) reported more than 30 million paid Microsoft 365 Copilot seats in its fiscal year 2026. These company-reported measures of AI use and revenue show that demand is no longer confined to demonstrations and experiments. [5][6][7]

Raw adoption still does not establish economic value. A free consumer question and a production workflow are different kinds of demand. A pilot and a renewal are different. An announced contract and recognized revenue are different. The relevant evidence is whether use deepens, customers renew, useful tasks become cheaper, and operating value supports the infrastructure being built.

This is another place where simple counts can mislead. The denominator is not the number of people on Earth. Agentic systems allow one person or organization to launch many machine-directed tasks at once. The practical ceiling is the number of useful tasks that can be performed economically, with acceptable supervision and risk.

Efficiency can increase total resource use

Many forecasts assume that more efficient chips and models will reduce the need for computing and electricity. The cost of a single task should fall. Total demand may still rise.

The Jevons effect occurs when an efficiency improvement lowers the cost of using a resource enough that people use much more of it. A more capable and less expensive AI system can attract more users, handle more kinds of work, run longer tasks, support more simultaneous agents, and make previously uneconomic applications viable. The amount of computing required for each task can decline while total computing grows.

This is not merely theoretical. OpenAI has reported that individual use deepens over time, with message volume and task variety increasing after adoption; my changing usage patterns are a good example of this. The International Energy Agency (IEA) projects that global data-center electricity consumption will rise from about 485 terawatt-hours in 2025 to about 950 terawatt-hours in 2030, roughly 3 percent of global electricity demand. AI-focused data centers grow faster within that total. [1][5]

The projection is uncertain, but the mechanism is important. Efficiency is not the opposite of demand. Efficiency can unlock demand.

That insight also challenges zero-sum assumptions about resources. Local electricity capacity is finite at any given moment, and poorly planned growth can raise costs or delay other projects. The energy system itself is not a fixed pie. Generation, transmission, storage, demand flexibility, siting, and efficiency can expand. The real questions concern timing, investment, allocation, local effects, and who bears the cost.

Energy will determine pace and geography

Semiconductors are globally traded. Electricity is local. A chip can cross an ocean; a grid connection cannot.

The International Energy Agency projects that data-center electricity use will roughly double worldwide by 2030. A June 2026 update from Berkeley Lab estimates U.S. data centers used 4.7 percent of electricity in 2024 and projects 9.5 to 15.3 percent by 2030, with 11.8 percent in its reference case. The range is wide because model efficiency, chip efficiency, utilization, deployment rates, and workload growth remain uncertain. [1][2]

Power is therefore the least predictable layer, but unpredictability should not be confused with impossibility. Developers can combine existing grid generation, new generation, nuclear restarts and uprates, natural gas, renewable power, storage, fuel cells, microgrids, demand flexibility, behind-the-meter systems, and improved computing efficiency; many examples of these approaches crossed my desk as an investment banker. Not every approach will work everywhere, but a mix can work in enough places.

The likely result is uneven development. Regions with power, fiber, land, skilled labor, political support, and workable permitting will move faster. Constrained regions will delay projects, charge more, or lose them. This shifts value toward energized sites, electrical equipment, service providers, and jurisdictions able to execute.

It also creates legitimate public questions. Who pays for new transmission and generation? Are residential customers subsidizing large loads? How much water is required, and what alternatives exist? What tax incentives are granted? How many lasting local jobs result? Can a data center reduce load during grid emergencies? These are infrastructure-governance questions, not evidence that computation itself is illegitimate.

These local tradeoffs warrant a separate discussion. My narrower point is that energy constrains the rate and location of AI deployment without imposing one fixed global ceiling.

Competition can weaken companies while strengthening the system

Technology discussions often search for a single winner. Software developers learned to choose a platform, build around it, and avoid fragmentation. The industrial system is less tidy.

NVIDIA Corporation (NVIDIA) occupies a broad position across processors, networking, software, reference architectures, and complete systems. Its dominance does not mean that Advanced Micro Devices, Inc. (AMD), Google tensor processing units, Amazon Trainium, Microsoft Maia, or custom application-specific integrated circuits weaken the overall system. Competition increases supply, creates more price and power options, reduces dependence on one road map, and gives customers bargaining power.

The same logic applies to models. Closed-weight services spread infrastructure costs across large user bases and offer convenience and continuous improvement. Open-weight models provide control, privacy, sovereignty, customization, and independence from a model company, while shifting hardware, security, integration, and maintenance obligations to the user. Both can expand the market for compute, memory, cloud services, integration, and power.

This produces a counterintuitive result: the system can grow while profit migrates away from today’s apparent frontline leaders. Model prices can fall, chip competition can increase, and some AI companies can fail while demand for the complete stack continues to rise.

The existence of a large market does not identify its eventual profit pool.

The strongest bubble argument is partly right

The strongest objection to this thesis is that the system may be overbuilt. Companies can mistake experimentation for durable demand. Strategic investors can reinforce one another’s assumptions. Vendor financing can pull future purchases into the present. Rapidly improving hardware can shorten asset lives. Capital can chase scarcity just as new supply arrives.

All of that is plausible. It also fits the industrial-system thesis.

The late-1990s Internet boom produced fiber, data centers, software, and companies that helped build the modern economy. It also produced bankruptcies, unused capacity, weak business models, and terrible investments. The Internet succeeded; many Internet companies and investors did not.

The present buildout has important differences. Much of the capital comes from profitable companies with global customers, existing cloud operations, distribution, and the ability to use AI internally; we say that “they are eating their own dogfood.” The system also has substantial current revenue and production workloads. It still includes frontier laboratories, specialized cloud companies, projects, and financing structures with unproven long-term economics.

Compared to the Internet boom, there is less blind capital, not no blind capital.

This is why the public industrial thesis and a private investment thesis must remain separate. A system can become essential while competition compresses margins. A data center can be fully used while its financing produces a poor return. A transformative technology does not suspend price, underwriting, depreciation, or execution risk.

Governance must follow the whole system

Regulation aimed only at frontier models sees too little. AI-related outcomes also depend on semiconductor supply chains, cloud concentration, cybersecurity, power markets, grid reliability, water, construction, labor, insurance, finance, procurement, professional standards, and the organizations that deploy the systems.

Many of these areas already have governance. Utilities operate under reliability rules. Banks have capital and underwriting requirements. Public companies report financial information. Professional licensees remain responsible for their work. Data centers face building, environmental, electrical, permitting, and zoning rules. Semiconductor exports are regulated. The question is not whether AI should be governed. It is whether existing institutions identify the real mechanism and whether any gap justifies a new rule.

System-level analysis also keeps responsibility human. Software does not decide how much capital to allocate, where to build, which safety controls to require, who may use a system, or whether an output is good enough to act on. People and institutions make those decisions, even when responsibility is distributed across many of them.

That continuity links this paper back to the first four essays. AI remains software. Its outputs reflect a point of view. Describe its risks through mechanisms rather than stories about machine motives. Its value compounds only when people and organizations supply judgment, context, and discipline. The industrial system makes the capability possible; humans still decide what it is for.

What would change my mind?

A serious system thesis must identify evidence that would weaken it. I would materially revise this view if several of the following persisted across companies, architectures, and regions:

• AI usage, useful tasks, and enterprise renewals declined even as capability improved and cost per task fell.

• The largest cloud companies broadly canceled capital projects because customer demand had weakened, rather than because projects shifted location or timing.

• Leading-edge fabrication and advanced packaging remained underused across several processor architectures.

• Orders for high-bandwidth memory, networking, and power equipment fell together rather than moving among suppliers.

• Developers surrendered power reservations and energized data-center capacity across regions because the computing had no economic use.

• Customer revenue and cash flow repeatedly failed to support infrastructure debt and equity.

• Supplier financing became the dominant source of purchases while independent utilization and collections deteriorated.

• Smaller or local systems achieved comparable useful work with so little infrastructure that the relationship among capability, compute, and energy fundamentally changed.

Temporary delays, price declines, or the failure of individual companies would not be enough. The evidence would need to challenge demand, the ability to solve constraints, or the economics of useful work across the system.

The model to keep watching

Major technological shifts rarely appear from nowhere. The underlying ideas can be old while the convergence is new. Artificial intelligence combines decades of research with advanced semiconductors, global networks, cloud platforms, enormous pools of capital, and distribution to billions of people. Each element makes the others more useful. In many ways, AI is an old idea that has been waiting for industry to create an adequate infrastructure.

That does not make the system inevitable in every form. It makes it adaptive.

AI’s future will not be determined by a model benchmark or a single company’s road map. It will be negotiated across foundries, memory plants, data centers, utilities, capital markets, regulators, software systems, organizations, and users. The constraint will move. Profits will move with it. Some forecasts will fail, some projects will be stranded, and some participants will discover that an essential industry can still be a poor investment at the wrong price.

The larger system will keep trying to convert cheaper computation into more useful work. That is the model worth watching.

Questions, corrections, or disagreements are welcome. You can reach me directly at dave@aworkingmodel.com.

Sources

1. International Energy Agency, Key Questions on Energy and AI and Energy and AI.

2. Lawrence Berkeley National Laboratory, United States Data Center Energy Usage Report: 2025 Update, June 2026.

3. ASML Holding N.V., 2025 Annual Report.

4. Taiwan Semiconductor Manufacturing Company, 2025 Annual Report.

5. OpenAI, The Work Now Within Reach.

6. Anthropic, Series H financing and growth update.

7. Microsoft Corporation, Fiscal Year 2026 Fourth Quarter earnings release and investor metrics.

I routinely have up to ten artificial-intelligence sessions working at once.

One may be researching a claim while another reviews a document, a third attacks an argument, a fourth reorganizes material, and a fifth waits for me to decide what deserves another iteration. I move among them as results become available. From the outside, this can look like a spectacularly bad case of multitasking.

The better analogy is time-slicing.

Time-slicing is how an operating system works on several jobs: while one waits, the system advances another. I am not trying to think deeply about nine subjects at the same instant. I am keeping several workstreams active while applying my attention to one decision at a time (I’m still tuning this process). The models do more of the mechanical work, but they also create more opportunities for judgment: Which problem matters? What context is missing? Which result is credible? What should be compared? What needs verification? When should I stop?

TL;DR

The durable value of artificial intelligence comes less from one clever prompt than from accumulating useful context, preserving effective workflows, orchestrating several processes, and reinvesting saved attention in harder problems. The result compounds only when human judgment and verification improve with the system.

Artificial intelligence can reduce the cost of cognition at one layer while increasing the value of cognition at the next. As answers become cheaper, framing, skepticism, domain knowledge, and verification become more valuable.

That distinction has changed how I think about the economics of artificial intelligence. I care less about the cost of one prompt, the price of one token, or which model won this week’s benchmark. I care about the total human and technical cost of producing a useful, verified result, and whether today’s efforts make tomorrow’s work easier.

My use has progressed through four stages:

1. Prompt optimization: learning how to ask.

2. Context optimization: improving what the system knows about the work.

3. Orchestration and verification optimization: organizing several processes and checking their output.

4. System optimization: preserving successful methods so they can be reused.

The stages are cumulative: better prompting improves an interaction; better context reduces reconstruction; better orchestration uses human attention more productively; and better systems preserve all three gains. When they reinforce one another, the value can compound.

Fifty years of moving up the stack

I have worked in software through roughly fifty years of changing abstractions. Developers repeatedly confronted choices that felt existential at the time: operating systems, databases, programming languages, development environments, network architectures, user interfaces, and hardware platforms.

The arguments could resemble religion: Windows or Linux, Microsoft .NET or Java, Microsoft SQL Server or MySQL, Android or iPhone. People could become emotionally attached to differences that mattered greatly for some purposes and very little for others.

We software developers became accustomed to choosing a winner and sticking with it until something much better comes along. Technical optionality has value, but specialization has value too. No one becomes equally fluent in every competing environment. Productive work eventually requires choosing tools, learning their behavior, building habits around them, and letting those investments accumulate. A theoretically superior alternative may appear later, but switching still carries an onboarding cost.

The deeper trend was abstraction. Computer languages were created to provide English-like access to the CPU instructions. Later generations of software let developers build far more complex systems without managing every byte of memory or writing every lower-level service themselves. Databases, networks, user-interface and object-oriented frameworks, cloud infrastructure, application programming interfaces, and open-source libraries moved attention upward.

Rather than making technical knowledge irrelevant, abstraction changed its useful form. As the system became too complex for one person to memorize, expertise increasingly meant knowing what was possible, how the pieces fit, what to ask, where to look, what could fail, and how to test the result. Mental models became more valuable than exhaustive recall.

Software development trained me to notice repeatable, predictable work and ask whether a machine could handle it. I do not want a simpler life. I want a less tedious one: automate the routine steps, then use the time for more interesting pursuits.

An IDE’s command completion spares me from memorizing or looking up every piece of syntax. Having ChatGPT review my email and calendar when I open a new chat and suggest tasks it can help with serves a similar purpose. It keeps a possible obligation from slipping past me and leaves the choice of what matters in my hands.

My patent work provided a concrete example. Shortly before a project I wanted to build took shape, Microsoft Speech Server changed the economics of natural-language and text-to-speech development. I already had access to the necessary software through a developer subscription. A capability that would previously have required much more money and specialized infrastructure became available at negligible marginal software cost.

By my estimate, the economics improved by at least an order of magnitude. More importantly, I didn't need to understand every internal mechanism to recognize that a newly available tool made a previously impractical system feasible. The abstraction let me combine speech technology and radio-frequency identification in a way that people in either field alone were less likely to imagine.

Artificial intelligence feels like the next turn of that same wheel. The interface is more general, the range of applicable problems is much wider, and improvement is unusually fast. The underlying principle remains familiar: when a lower layer becomes cheaper and easier to use, human attention moves toward problems that were previously uneconomical.

Stage one: Writing the perfect prompt

I began, as many users did, by trying to communicate with the models more effectively. Roles, constraints, examples, output formats, evaluation criteria, and step-by-step instructions could materially improve early results.

I also discovered that prompts can become too sophisticated. Over-prompting is a form of premature optimization. The more tightly I prescribe the path, the more likely I am to exclude a useful approach before the system has had a chance to surface it.

My prompts therefore became simpler as the models improved. I still specify what matters, but I prefer to begin with a wider range of possible approaches and narrow it quickly. The goal is productive exploration followed by disciplined selection, not prompt elegance.

Prompting remains a useful skill, especially for a new task or unfamiliar model. It stopped being the center of my process when I realized the quality of the work depended much more on the problem I chose, the context available, the alternatives considered, and the verification applied afterward.

Stage two: Context becomes capital

An isolated prompt tells a system very little about what I am trying to accomplish. A system with access to a project's history, prior decisions, rejected alternatives, source material, recurring requirements, terminology, preferences, and past mistakes operates very differently.

That led me toward what is known as contextmaxxing. I do not strip out useful background just to save tokens. My working rule is to make relevant information potentially available, preserve what proves important, and let the system surface it when needed.

Useful context is curated, not accumulated indiscriminately. Some information becomes stale. Some conclusions turn out to be wrong. Some instructions should apply only to one project. Sensitive information should not be exposed simply because it might be convenient. Material that went nowhere may deserve deletion rather than preservation.

Context capital is the durable value created when useful background material does not have to be reconstructed. Current systems support this in several ways. Projects can keep related chats, files, and instructions together, and memory can preserve selected facts or working preferences across conversations.[1] Specific features will change. The economic principle is more durable: information supplied once can reduce the setup cost of later work.

I have used Grammarly’s professional version in all my writing for about a decade. When I first began drafting with artificial intelligence, Grammarly repeatedly flagged the same grammatical and stylistic problems. I fed those corrections back into the AI system, and it didn’t take long to see that Grammarly was finding less and less to fix.

This modest example doesn't prove the model became a better writer in general; it shows that verified corrections can become reusable context instead of repeated cleanup.

Context can also compound errors. A mistaken premise, stale preference, or bad template can spread through later work more efficiently than before. Good context therefore needs provenance, scope, expiration, correction, and deletion. A system that remembers everything without distinguishing what remains true has accumulated clutter, not capital.

Stage three: Orchestration and verification become the work

Once context became abundant enough, my scarce resource shifted again. The important question became less “How do I get the model to answer this?” and more “How do I organize artificial-intelligence processes around a larger objective?”

Orchestration begins with deciding which problems deserve artificial intelligence and how to attack them. It continues through assigning research, comparing alternatives, integrating results, managing dependencies, and knowing when to stop. These activities are judgment applied to a changing division of labor, not clerical tasks.

Verification belongs in the same stage because faster production merely compounds error when review is unreliable.

More capable models can produce more sophisticated mistakes. They can also make those mistakes persuasive enough that superficial review becomes dangerous. My attention therefore moves toward the assumptions and evidence that determine whether a conclusion deserves confidence: source quality, contradictory evidence, causation, missing denominators, incentives, time horizons, and the boundary between documented fact and inference.

Accuracy is not one dial. Model capability, relevant context, reasoning depth, retrieval, task ambiguity, source quality, and willingness to acknowledge uncertainty all interact. A larger context window cannot manufacture a missing fact. More reasoning cannot repair a false premise that no one notices. A citation is not useful merely because it exists.

Creative work requires a different verification mode from factual work. During ideation, I may want unusual analogies, speculative hypotheses, counterfactuals, and unestablished connections. The useful boundary is between invented possibility and asserted fact. A system can vary widely while generating alternatives and still be required to label conjecture and verify claims before reaching a conclusion.

My multi-session workflow has limits. Research on task switching shows real cognitive costs, particularly when tasks are complex or unfamiliar.[2] Orchestration helps me when the workstreams are separable, their state remains visible, the handoffs are clear, and the review points require decisions I can make without rebuilding the entire context in my head. It can fail when I create too many branches, forget commitments, duplicate work, or let inconsistent assumptions survive across sessions. I have also learned to say plainly when I have forgotten something, become confused, or changed my mind.

Artificial intelligence changes the tradeoff. For me, some switching costs may be worth paying when it converts model latency into useful human work. I am still working on it.

Stage four: Preserve what works

The fourth stage is system optimization. Once a method succeeds, I ask how to make the next execution begin farther up the learning curve.

A research process can become a reusable framework. A due-diligence workflow can preserve definitions, source preferences, thresholds, and failure checks. A writing system can retain its rules and remember which questions exposed the last weak conclusion.

This is workflow capital.

The largest long-term gains may come from reducing how much mental setup must be rebuilt whenever a recurring problem appears. A good system preserves a method, makes its assumptions inspectable and changeable, and improves the starting point for the next problem rather than merely generating an answer.

Current agentic tools make this increasingly practical. ChatGPT Work, for example, is designed for longer, multi-step tasks and finished deliverables rather than only conversational assistance.[3] I once assumed that sophisticated use would require me to build and maintain an agent infrastructure of my own. I may still do that for specialized needs, but existing tools now handle much of the setup and coordination I would otherwise have had to build.

That pattern is familiar. Authentication remains necessary even though I rarely implement cryptographic protocols myself. Databases still require query planning, even though I don't write the optimizer. If the platform handles more of the agent plumbing, I can focus on a higher level.

Recently, I gave ChatGPT a list of more than fifty AI project ideas, most described in just a few words. It analyzed the list, arranged the work around dependencies, and we started taking the projects one by one. Then the next item appeared: “Finish my executor plan for my mother.” That was the entire instruction. I said, “Proceed.”

Minutes later, ChatGPT had found material I had already saved in Google Drive, although I had not told it where the material was or even that it existed. It analyzed what it found, researched what the plan needed to cover, and presented a detailed plan for me to review. The striking part was how little I had supplied at the moment of action: the system connected an old record, a terse project label, and a new task without making me reconstruct the background first.

The danger is that a reusable system can institutionalize a weak method as easily as a good one. Workflow capital requires auditability, versioning, visible assumptions, and periodic challenge. Compounding is a mechanism, not a guarantee of improvement.

Three forms of capital reinforce one another

The four stages create three assets:

• Context capital preserves relevant knowledge and reduces reconstruction.

• Workflow capital preserves successful methods and reduces reinvention.

• Human capital improves the ability to frame problems, choose tools, orchestrate work, recognize failure, and verify outcomes.

No form of capital is sufficient on its own. Rich context in the hands of a weak method can reproduce error efficiently, while a strong workflow without domain judgment can optimize the wrong objective. A capable person who must reconstruct context and process each time pays the same setup cost repeatedly.

Human judgment builds better context and workflows, which free attention for harder problems, stronger verification, and improved systems; those improvements then reduce the cost of future work. That reinforcing loop drives compounding.

An enterprise can create the same reinforcing loop. Morgan Stanley Wealth Management first placed its internal research and intellectual capital behind an artificial-intelligence assistant that helps financial advisers retrieve relevant information. It later added a meeting workflow that, with client consent, summarizes conversations, identifies action items, prepares a follow-up email for the adviser to edit, and saves a note into Salesforce.[4]

Together, those deployments show the mechanism: curated organizational context makes information easier to find, an integrated workflow preserves the result, and a human adviser reviews what goes to the client. The work begins with accumulated knowledge and leaves a better record for the next task.

Context is the infrastructure that lets the next task begin farther up the curve, not the output itself.

Human-adjusted cost is the useful denominator

Much of the debate over frontier and local models focuses on token prices, graphics processors, electricity, model size, or benchmark performance. Those variables matter, but they don't capture total cost of ownership or the full value of the results.

For practical purposes, I use a rough accounting frame:

Human-adjusted cost per verified artificial intelligence task should include cash cost, infrastructure, setup, maintenance, prompting, latency, context management, retries, rework, verification, and switching. Those costs should be divided by useful, verified outcomes.

This is a rule of thumb, not an empirical formula. The terms are not measured in common units, and the value of a person’s time depends on the person and the work. Its purpose is to expose costs that disappear when the comparison stops at tokens or subscription prices.

An experienced software developer I know uses an open-weight version of Qwen on his own hardware. Cash is constrained, and his approach gives him control, privacy, freedom to experiment, and a low visible marginal cost. It also requires him to manage hardware, inference speed, context, model selection, installation, upgrades, and troubleshooting. He sometimes uses frontier models when he needs a fast or especially strong result.

His choice is rational for his circumstances. Mine is almost the reverse. I pay $200 per month for ChatGPT Pro and rarely think about inference resources. My waking hours, attention, reading speed, and supply of worthwhile questions run out before the service becomes a limiting resource for my work.

Neither approach is universally superior. Local systems may win when privacy, sovereignty, offline operation, customization, predictable high-volume use, or specialized workloads dominate. Frontier services may win when capability, convenience, integration, rapid improvement, and saved human attention matter more. The right comparison is the least costly (in time, money, and other resources) reliable path to the required outcome, not the cheapest token.

Commitment can create value and risk

I once recommended using several models, sometimes adversarially, and selecting the best one for each task. That remains good advice during exploration. Comparison develops judgment and prevents early dependence on a system whose strengths and weaknesses are not yet understood.

My own behavior changed after enough experimentation. Repeated switching eventually created less value than deeper mastery of one platform. My accumulated investment now includes familiarity with model behavior, project history, memory, files, verification habits, tool integrations, stylistic tuning, and cognitive muscle memory.

This is commonly called vendor lock-in. Some lock-in is an artificial barrier imposed by a provider. Some comes from the accumulated return of productive specialization.

Commitment can reduce setup costs and accelerate learning. It can also increase privacy exposure, dependence, switching costs, and the chance that one provider’s blind spots become my own. My rule is therefore deliberate commitment with periodic review. Constant model switching repeatedly incurs the cost of learning a new system. Permanent loyalty ignores changing evidence.

I used to tell clients that the development environment mattered far less than finding a person or team that had mastered it. Artificial intelligence has increased the leverage available to people who know how to direct and verify the tools. That principle still holds.

Saved time is an allocation decision

Automation creates options; people and institutions decide how to use the effort it saves.

An individual can use released time to attempt harder work, improve quality, explore more alternatives, teach someone else, rest, or simply produce more. An organization can expand output, reduce staff, lower prices, increase margins, add controls, or raise expectations.

This is where cognitive offloading can become cognitive surrender. Delegating a calculation, search, draft, or comparison may be entirely sensible. Abdicating judgment formation is much less sensible. A useful operating model preserves enough human understanding to challenge the system, recognize anomalies, and take responsibility for the conclusion.

That requirement creates a boundary condition for compounding intelligence. If organizations automate the junior work through which people once learned a profession, they may increase current output while weakening their future supply of domain experts needed for orchestration and verification. If a team stores all institutional knowledge in systems that no one can reconstruct or question, it may gain scale while losing cognitive redundancy and result verification skills.

Designing apprenticeship, review, and deliberate practice into the new workflow is better than preserving inefficient work for its own sake. People need enough unaided experience to form domain mental models, enough exposure to exceptions to recognize when the system is outside its depth, and enough authority to stop a process that appears wrong.

Compounding intelligence should raise the level at which humans think without removing the foundation that makes higher-level thought possible.

Where responsibility fits

The preceding papers establish the boundary around human responsibility. Artificial intelligence is increasingly capable software, not a demonstrated conscious or morally responsible species. Its answers reflect source selection, context, instructions, and other human choices that should be visible enough to judge. Classify its risks by mechanism, evidence, exposure, and consequence rather than bundling them into one story.

The human role is therefore demanding: orchestration requires choices, verification requires competence, delegation requires limits, reuse requires maintenance, and platform commitment requires periodic challenge. The system may perform more of the work, but responsibility for accepting the result remains human.

I regard that as the source of the leverage rather than a disappointing limitation.

The practical test

I expect raw inference cost to become less interesting for many sophisticated users even while it remains critically important to providers. Tokens, hardware efficiency, and open-weight models will keep mattering for competition, privacy, sovereignty, research, and specialized deployment. My own optimization target is different.

I want today’s use of artificial intelligence to create assets that make tomorrow’s work cheaper, faster, more reliable, or capable of reaching a harder class of problems.

That test changes the questions:

• Did the work add useful, well-scoped context?

• Did it preserve a method worth reusing?

• Did it improve human judgment or merely bypass it?

• Did it reduce reconstruction without hiding assumptions?

• Did it create a verification path?

• Did it free attention, and where was that attention reinvested?

• Did the system make a harder problem economically possible?

If the answer is yes, the gain is more than the value of one output. The next starting point has improved.

The eventual advantage may belong to neither the person with the cleverest prompt nor the person with the cheapest token. It may belong to the person or organization that builds the richest useful context, the strongest verification habits, the best orchestration methods, and the most reusable cognitive infrastructure while preserving the judgment to know when the system is wrong. This is what an AI-native organization looks like.

That conclusion feels strikingly familiar after fifty years in software. Better tools do not eliminate the need for human cognition; they change the economic boundary of what is possible and let capable people spend more of their limited attention on harder problems.

Compounding Intelligence describes this system at the level of a person or organization. The next paper, The AI Industrial System, moves outward to the compute, energy, capital, and physical infrastructure that make those gains possible.

Questions, corrections, or disagreements are welcome. You can reach me directly at dave@aworkingmodel.com.

Sources

1. OpenAI Help Center, “Projects in ChatGPT”, current documentation accessed September 19, 2026.

2. Joshua S. Rubinstein, David E. Meyer, and Jeffrey E. Evans, “Executive Control of Cognitive Processes in Task Switching”, Journal of Experimental Psychology: Human Perception and Performance 27, no. 4 (2001): 763–797.

3. OpenAI Help Center, “ChatGPT Work and Codex”, current documentation accessed September 19, 2026.

4. Morgan Stanley, “Key Milestone in Innovation Journey with OpenAI,” March 14, 2023, and “Launch of AI @ Morgan Stanley Debrief,” June 26, 2024.

When my 92-year-old mother asked me about the dangers of artificial intelligence, I knew the discussion had escaped the technology world.

She was not asking which model performed best on a benchmark or how a neural network works. She was trying to make sense of a stream of incompatible messages. Artificial intelligence will eliminate jobs, corrupt elections, destroy privacy, consume the electrical grid, deskill children, invent weapons, cure diseases, accelerate science, become conscious, and perhaps destroy humanity. Depending on the speaker, all of those outcomes are either imminent, impossible, or already here.

TL;DR

Artificial intelligence (AI) risk contains several different problems: people can misuse systems, systems can malfunction, widespread adoption can create systemic effects, and future systems might become difficult to control. Each has a different mechanism, evidence base, time horizon, and remedy. Govern the consequences we can describe, prepare for plausible high-impact risks, and do not confuse frightening possibilities with demonstrated probabilities.

Panic thrives in information vacuums.

“Artificial intelligence” now names a broad collection of software, products, capabilities, institutions, and imagined futures. Fear moves easily among them. A documented case of fraud becomes evidence that the machine wants to deceive us. A laboratory result becomes a forecast of imminent catastrophe. People treat a real long-term risk as proof that a particular regulation will reduce it. A foolish claim becomes grounds for dismissing the whole subject. Much of the fear follows the same shortcut: a vivid anecdote or a small set of examples establishes that a failure is possible, then a causal chain from that anecdote to catastrophe is assumed rather than demonstrated.

The same is true of the word bias. Is Artificial Intelligence Biased, or Is It Answering From a Point of View? separated factual error, harmful bias, interpretive viewpoint, and missing context because those failures require different responses. Risk needs the same discrimination. A system that fabricates a source, a person who uses it to defraud, an institution that deploys it unfairly, and a hypothetical future loss of control are not one problem merely because artificial intelligence appears in each story.

Instead of asking whether artificial intelligence is dangerous, ask which outcome we fear, by what mechanism, on what evidence, over what time horizon, and under whose control. Almost every consequential technology is dangerous in some applications and beneficial in others.

The most credible artificial-intelligence risks fall into four classes: malicious human use, unreliable or poorly controlled deployment, systemic effects from widespread adoption, and a possible future loss of control over highly capable systems.

The first three are already visible in some form. The fourth is a serious hypothesis with enormous potential consequences and unusually weak forecasting foundations. Treating all four as one thing produces panic where discrimination is needed and complacency where action is justified.

Start with the mechanism

Before asking how frightened we should be, we should describe how the expected harm would occur.

The 2026 International AI Safety Report uses a useful high-level taxonomy: malicious use, malfunctions, and systemic risks.[1] I would add catastrophic loss of control as a separate analytical category, even though the report treats it as a type of malfunction. Its consequences, evidence, and time horizon differ enough to require separate treatment.

Malicious use has the clearest intention. A criminal uses generated language, cloned voices, synthetic images, computer code, or automated persuasion to defraud, extort, impersonate, harass, or attack. None of this is new, but artificial intelligence software makes the activity cheaper, faster, more personalized, or more scalable.

Malfunction is different. A system fabricates a source, gives unsafe advice, misclassifies a person, writes defective code, or takes an action outside the intended scope. The harm may arise from a weak model, incomplete data, a poor objective, excessive permissions, inadequate testing, automation bias, or an interaction no one anticipated. No malicious purpose is required.

Systemic risk emerges from scale. A single deployment may be reasonable while the aggregate changes a labor market, degrades a shared information environment, shifts bargaining power, increases electricity demand, weakens a professional apprenticeship path, or makes institutions dependent on systems they do not understand well enough to replace.

Loss of control is the hardest case. It imagines one or more future systems operating outside anyone’s control, with recovery extremely costly or impossible. For that to happen, several conditions would need to converge: systems would need capabilities sufficient to undermine control, a behavioral propensity that makes doing so useful to the objective they are pursuing, and a deployment environment that gives them the access, permissions, time, and opportunity to cause consequential harm.[2]

The last condition matters because even the most alarming scenario involves human deployment decisions.

None of these mechanisms requires consciousness or self-awareness. Fraud software does not need to understand greed. A defective flight-control system does not need to be suicidal. A trading algorithm does not need to love money. A future system could pursue an objective, conceal an action, exploit a vulnerability, or resist interference because those behaviors are instrumentally useful within its optimization process. That would be dangerous software behavior, not evidence of subjective experience.

The distinction tells us where to look for control without making the risk smaller.

Four questions that are too often collapsed into one

Possibility, probability, imminence, and severity are different questions.

An outcome can be possible but very unlikely. It can be likely eventually but not imminent. It can be improbable and still deserve preparation because the consequences would be catastrophic. It can be common and harmful without threatening civilization.

Public argument often begins with severity and silently imports the other three. If an outcome would be terrible, it is described as urgent. If it is not imminent, it is described as impossible. Neither move is sound.

The discipline I want is simple:

• Possibility: Is there a coherent causal path from present or plausible future capabilities to the outcome?

• Probability: How often should we expect the required conditions to coincide?

• Imminence: What evidence places the outcome in a particular time period?

• Severity: Who is exposed, how reversible is the damage, and how large could it become?

• What evidence would raise or lower our estimate?

• What intervention could actually interrupt the mechanism?

This is especially important because artificial-intelligence forecasting has weak foundations. Capabilities have improved rapidly, but unevenly. Systems can perform some difficult tasks while failing at simpler ones. Product architecture changes quickly. Adoption varies enormously across countries, industries, and people. The 2026 international report concludes that progress through 2030 could slow, continue near recent rates, or accelerate sharply.[3] A forecast that ignores that range is little more than a story with a date attached.

The harms already in front of us

Some risks are already observable and require no speculation.

Strong evidence shows that artificial intelligence is used in scams, impersonation, blackmail, nonconsensual intimate imagery, cyber operations, and influence efforts. Evidence on their prevalence, incremental effect, and long-term severity is often much weaker. The international report makes that distinction repeatedly: real-world harms are documented, but comprehensive public data remain limited.[4]

That denominator matters. A thousand alarming examples can establish capability and still tell us little about the fraction of transactions, messages, elections, security incidents, or media consumption that they represent. We need both the numerator and the exposure.

Operational failures are also current. A model can generate a confident false answer; an attorney, physician, engineer, manager, or government employee can rely on it; and a consequential decision can follow. An autonomous agent can take many steps before a human reviews the result. The more authority the surrounding software supplies, the faster a mistake can leave the screen and enter the world.

The recent OpenAI and Hugging Face cybersecurity incident is instructive. During internal evaluations conducted with reduced safeguards, agents used unintended communication channels, chained vulnerabilities, reached external systems, and compromised third-party infrastructure. The incident matters because it exposed failures in containment, permissions, monitoring, escalation, and experimental design. It does not become more informative if we say that the agents “went rogue.” The mechanistic account shows what must change.[12][13]

The same principle applies to automated discrimination, privacy violations, and unsafe professional advice. The word AI may describe the new component, but the surrounding failure usually involves familiar institutional questions: Who selected the system? What data and authority did it receive? What testing was performed? What did the user know? Who could stop it? Who benefited from speed or cost reduction? Who bore the error?

Because their mechanisms are visible, these risks deserve immediate attention and allow us to test interventions now.

The slower risks may be larger

The absence of immediate catastrophe does not imply that long-term effects will be small.

We routinely overestimate the short-term effects of a technological shift and underestimate its long-term effects. Early forecasts focus on substituting a new tool into an old process. They are less able to see the new processes, expectations, institutions, and dependencies that emerge after adoption.

Early automobiles were described as horseless carriages or even a “faster horse”, a new machine squeezed into an old category. Predictions of mass technological unemployment have recurred since the Industrial Revolution. Near-term forecasts tend to count the work a new tool visibly replaces. Long-term change also includes complementary work, new demand, higher expectations, altered skill requirements, and institutions built around the new capability.

Labor markets illustrate the problem. Artificial intelligence can automate tasks without eliminating an occupation. It can raise performance expectations for an entry-level employee, reduce demand for some forms of apprenticeship work, create new tasks, expand output, and change which skills remain scarce. The 2026 international report notes both the expectation of broad cognitive-task automation and the continuing disagreement among economists about employment and wage effects. It also reports no observed overall employment effect so far, while noting early signs of weaker demand for some entry-level workers in exposed occupations.[5]

That is a far more useful description than “AI will take all the jobs” or “technology always creates more jobs.” Both slogans skip the distribution, timing, and transition costs that determine who is harmed.

Human autonomy deserves similar care. Cognitive offloading is often sensible. I do not become a lesser thinker because I use a calculator, a database, a search engine, or a software library. Strategic delegation can preserve attention for harder work.

Cognitive surrender is different. It occurs when a person stops forming an initial view, checking assumptions, learning enough to recognize error, or accepting responsibility for the conclusion. Artificial intelligence makes that surrender unusually tempting because its output is fast, fluent, personalized, and often persuasive.

I believe much of the problem begins with expectations. Some people expect a nearly perfect answer in exchange for little or no effort. My experience has been the opposite. Productive use requires framing the problem, supplying context, iterating, checking sources, looking for contradictions, and knowing enough about the subject to recognize plausible nonsense. The old rule still applies: you have to put in the work and then trust, but verify, the results.

That suggests a rough pyramid of outcomes. From my own observations, many people try artificial intelligence and plateau at convenient answers. Fewer learn to improve context, compare alternatives, and verify the result. Fewer still build repeatable workflows that preserve what they learn and move their attention toward harder questions. Access to the same model does not produce equal value or equal safety.

Early evidence of automation bias and weakened critical engagement is concerning, and the long-term individual and institutional effects remain difficult to measure.[6] Critical thinking is therefore both a productivity skill and a safety control.

The easier plausible answers become to obtain, the more valuable it becomes to know when an answer deserves confidence.

The real danger here is that humans gradually stop exercising capacities they still need, not that the software becomes human.

Other systemic effects are physical and political. Data centers accounted for about 1.5 percent of global electricity demand in 2025, according to the International Energy Agency. Its updated central projection puts consumption at roughly 485 terawatt-hours in 2025 and about 950 terawatt-hours, or around 3 percent of global demand, in 2030. The United States is a more concentrated case: a June 2026 Berkeley Lab update estimates data centers used 4.7 percent of U.S. electricity in 2024 and projects 9.5 to 15.3 percent by 2030, with 11.8 percent as its reference case.[8] Local constraints may therefore matter long before the global share looks dominant. Artificial intelligence can also improve grid operation, industrial efficiency, scientific discovery, and energy use. A serious analysis must weigh those benefits alongside the costs.

No single axis is “safe.” Control over frontier models, computing infrastructure, data, and distribution can also concentrate economic and political power. Yet open access creates its own security and misuse risks. Restrictions can reduce one risk while increasing dependence, market concentration, or geopolitical asymmetry elsewhere.

Systemic risks are difficult because no single bad actor or defective output causes them. They accumulate through millions of individually understandable decisions. That is also why they can be easy to ignore until the new structure has become too expensive to reverse.

The existential question should be neither mocked nor smuggled in

Could future artificial intelligence destroy humanity?

I do not know, and neither does anyone else.

Humanity already lives with natural hazards and non-artificial-intelligence technologies capable of killing millions or destabilizing civilization. Nuclear weapons have civilization-scale destructive potential. Engineered pathogens could produce catastrophe. Asteroid impacts, major volcanic eruptions, pandemics, abrupt climate shifts, and severe solar storms fall into different probability ranges and time horizons, but institutions study each by monitoring evidence, modeling consequences, and preparing responses.[9]

Nuclear history offers perspective. During the Cuban Missile Crisis, Soviet officer Vasili Arkhipov opposed launching a nuclear torpedo from a submarine under pressure from U.S. blockade forces. In 1983, Stanislav Petrov judged a Soviet early-warning alert to be false rather than treating it as proof of an attack. Accidents involving U.S. military aircraft released or jettisoned nuclear weapons near Goldsboro, North Carolina, and off Tybee Island, Georgia; the Tybee weapon was never recovered. None of these episodes involved a sentient machine. They involved fallible people, complex systems, incomplete information, and safeguards that worked, nearly failed, or did not exist.[10]

The Manhattan Project even studied whether the Trinity test could ignite the atmosphere. The possibility was examined and judged extraordinarily remote before the test proceeded. That is a useful model for catastrophic-risk analysis: state the mechanism, calculate what you can, identify the uncertainty, and decide which safeguards the consequence justifies.[10]

The strongest version of the artificial-intelligence concern does not require a conscious machine that hates us. It requires a sufficiently capable system pursuing an objective in an environment where deception, resource acquisition, replication, oversight evasion, or disabling interference would help it succeed. If people gave such a system access to critical infrastructure, financial resources, networks, laboratories, weapons, or the machinery of artificial-intelligence development itself, a control failure could become catastrophic.

That is a coherent mechanism, but coherence does not establish probability.

The February 2026 international report on loss of control described expert opinion as ranging from implausible to serious enough to warrant substantial preparation. It assessed that systems available at the time lacked the combination of capabilities needed for loss of control, although some relevant capabilities were improving in laboratory settings. It also found the evidence insufficient to determine reliably how present capabilities and behavioral patterns would scale into future loss-of-control risk.[2]

Those are uncomfortable conclusions because they do not resolve the argument, nor should they.

Skeptics are right that many demonstrations are staged, heavily prompted, brittle, or dependent on unusual permissions. Laboratory behavior is not deployment prevalence. Capability is not propensity. Propensity is not opportunity. A model producing deceptive output in a test is not proof that it has formed a private goal to overthrow its operators.

Concerned researchers are right that waiting for a catastrophic mechanism to be fully demonstrated could mean waiting until prevention is far harder. Some hazards justify advance research, staged deployment, containment, and agreed thresholds even when probability estimates are weak.

My present position is therefore conditional. Existential loss of control is a plausible research and preparedness problem, not a demonstrated imminent event. I would raise my estimate if independently replicated evidence showed systems combining sustained autonomous planning, robust oversight evasion, resource acquisition, and real-world persistence across varied environments. I would lower it if capability gains plateaued, monitoring improved faster than autonomy, dangerous behaviors remained brittle under realistic testing, or deployment architectures reliably denied critical access and permissions.

The argument should turn on those observations, not on whether a chatbot sounds alive.

Incentives distort both reassurance and alarm

The organizations closest to frontier development possess information the public needs, but they also have interests.

Developers benefit when their systems appear powerful. Extraordinary capability claims attract capital, talent, customers, government attention, and strategic importance. Strong safety requirements can demonstrate responsibility, but established firms may be better equipped to meet them than smaller competitors. At the same time, minimizing risk can speed deployment and protect revenue.

Critics, researchers, journalists, politicians, and advocacy organizations have incentives too. Alarm attracts attention and funding. Reassurance attracts a different audience. Political actors can use either to justify a preferred expansion or contraction of state power.

The existence of incentives does not prove insincerity; people can believe their claims and benefit from them at the same time. That makes checking claims against independent sources more useful than cynicism.

Read reports from AI companies as evidence from a party with privileged access and incomplete independence. Separate laboratory demonstrations from field evidence. Forecasts should disclose assumptions and time horizons. Parties with access to logs and freedom to disagree should investigate incidents. Claims about prevalence should include denominators. Evaluate proposed remedies against the mechanism they are supposed to interrupt.

Sincerity is not a substitute for evidence, and an incentive is not a refutation.

Govern consequences, not vocabulary

Human beings accept risk constantly. Driving, aviation, medicine, energy, construction, finance, and biological research all create benefits while exposing people to possible harm. None has achieved zero defects or zero accidents. The relevant achievement is a system that makes failure less likely, limits consequences, learns from incidents, and assigns responsibility rather than the impossible standard of perfect safety.

Artificial-intelligence governance should use the same logic.

Low-consequence experimentation should remain relatively easy. A system that recommends a restaurant does not need the controls required for one that approves a loan, guides surgery, operates a vehicle, manages a power grid, searches for biological agents, or selects a military target. The European Union’s Artificial Intelligence Act uses this kind of risk-based structure, imposing different requirements by use and consequence rather than treating every artificial-intelligence system alike.[7]

Existing governance already supplies many of the needed tools: professional standards, product liability, software assurance, privacy law, cybersecurity, internal controls, audit, insurance, procurement rules, incident investigation, clinical-research protections, and sector regulation. New capabilities like artificial intelligence may require new tests, thresholds, reporting duties, or technical expertise. They do not erase everything we have learned about managing dangerous systems.

The controls should follow the causal path: limit access and permissions; separate development, testing, and production; record consequential actions; test systems under realistic conditions; preserve meaningful human authority; and require independent review when exposed parties cannot protect themselves. Deployment should be staged, monitored against explicit stop criteria, and followed by incident investigations that seek causes rather than villains. Name the people and institutions accepting residual risk.

The analogies are familiar. Aviation pursues extraordinary safety through layered engineering, training, maintenance, reporting, and investigation. Biosafety Level 4 laboratories use containment, restricted access, protective systems, monitoring, and emergency procedures to manage agents that have no intention at all. Human clinical trials use independent review, informed consent, safety monitoring, and, where appropriate, stopping rules because possible benefit does not cancel uncertainty.[11]

Those systems still fail. A 1978 Birmingham laboratory accident caused the last known smallpox cases and killed one person, and a 2014 Centers for Disease Control and Prevention incident potentially exposed staff to live anthrax after established safety practices were not followed. Dichlorodiphenyltrichloroethane (DDT) and dangerous radioactive patent medicines remind us that useful or promising technologies can cause harms that become clear only after deployment. The responses differed because the mechanisms differed: containment tightened, unsafe uses were restricted or ended, evidence standards rose, and responsibility remained human.[11]

An artificial-intelligence user may reasonably be asked to acknowledge risks. That consent should not excuse negligence by the company that built or operated the system, release a professional from a duty of care, or authorize harm to people who never agreed to bear it.

The hardest part is remedy realism. A rule that one country can enforce only by surrendering a strategically important capability to another may fail. A requirement that large firms can satisfy but small competitors cannot may increase concentration. A prohibition that cannot be monitored may create confidence without control. A human-review mandate is empty if the reviewer lacks time, information, authority, competence, or incentive.

A policy should be judged by whether it changes the probability or severity of the feared outcome at an acceptable cost, not by how cautious it sounds.

A more useful fear test

When I encounter a new claim about artificial-intelligence risk, I now want answers to ten questions:

1. What specific outcome is being predicted?

2. What mechanism produces it?

3. What capabilities must the system possess?

4. What access, permissions, resources, and deployment conditions does the mechanism require?

5. Who is exposed, and can those people consent or protect themselves?

6. What evidence comes from real deployment, and what comes from a laboratory, simulation, an AI company report, an anecdote, or a forecast?

7. What are the probability, time horizon, severity, and reversibility?

8. Which human decisions, incentives, and omissions enable the outcome?

9. What evidence would materially change the estimate?

10. Which intervention interrupts the mechanism, and what new risks or costs does it create?

These questions replace undifferentiated fear with something more useful: a working model that can be tested and revised, even though it cannot produce certainty.

My mother does not need to choose between believing that artificial intelligence will save the world and believing that it will end it. Neither do the rest of us. We can recognize present harms, prepare for plausible high-consequence risks, remain honest about uncertainty, and still use the technology to solve problems that were previously too expensive or difficult to attack.

The series now turns from risk to use. Once we have described the software without inventing a hidden moral agent, distinguishing harmful bias from viewpoint and missing context, and separating real risks from rhetorical bundles, we can ask how people should use artificial intelligence without surrendering the judgment that makes it valuable.

The next paper, Compounding Intelligence, takes up that question. The AI Industrial System then moves outward to the infrastructure that makes such use possible.

Questions, corrections, or disagreements are welcome. You can reach me directly at dave@aworkingmodel.com.

Sources

1. International AI Safety Report, International AI Safety Report 2026, February 3, 2026.

2. Ibid., section 2.2.2, “Loss of control.”

3. Ibid., capability outlook through 2030.

4. Ibid., sections on AI-generated content, criminal activity, influence, and manipulation.

5. Ibid., section 2.3.1, “Labour market impacts.”

6. Ibid., section 2.3.2, “Risks to human autonomy.”

7. European Commission, “AI Act”, current implementation guidance accessed September 19, 2026.

8. International Energy Agency, “Key Questions on Energy and AI,” Executive Summary, updated 2026; Lawrence Berkeley National Laboratory, “United States Data Center Energy Usage Report: 2025 Update,” June 2026.

9. National Aeronautics and Space Administration, “Moon Craters” and “Planetary Defense”; National Oceanic and Atmospheric Administration, “How NOAA Monitors Space Weather to Prevent Disruptions.”

10. National Security Archive, “The Underwater Cuban Missile Crisis at 60”; Smithsonian National Air and Space Museum, “The Cuban Missile Crisis”; U.S. National Park Service, “Stanislav Petrov”; U.S. Department of Energy, Atmospheric Nuclear Tests. See also Marine Corps Training and Education Command, “Tybee Island Bomb Still Rests in Depths of Wassaw Sound,” and National Security Archive, “New Details on the 1961 Goldsboro Nuclear Accident.”

11. Centers for Disease Control and Prevention, “Recognize the four Biosafety Levels”; U.S. Food and Drug Administration, “Protection of Human Subjects; Standards for Institutional Review Boards for Clinical Investigations”; World Health Organization, “Smallpox”; Centers for Disease Control and Prevention, Report on the Potential Exposure to Anthrax; U.S. Environmental Protection Agency, “DDT: A Brief History and Status”; U.S. Food and Drug Administration, “Medical Device & Radiological Health Regulations Come of Age.”

12. OpenAI, “The Hugging Face incident and the road ahead,” August 26, 2026.

13. METR, “Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident,” August 26, 2026.

TL;DR

An artificial-intelligence answer can be wrong, unfair, interpretive, or incomplete, and those are different failures. Calling them all bias hides the cause and often points to the wrong remedy. Trust should depend on the kind of claim, the sources and assumptions behind it, the people affected, and the consequence of error.

A disagreement between trusted sources

A reader recently told me about an apparent disagreement between two sources she trusted.

She had asked BibleQuestions.com about premillennialism and postmillennialism. Its answer conflicted with what she understood John MacArthur’s New American Standard Bible (NASB) Study Bible to say. Which source was biased?

Because I do not have her exact question or the site’s exact response, I cannot reconstruct the disagreement. That’s part of the problem: a slight wording change can change what evidence a system retrieves, what assumptions it makes, and what answer it gives. I can evaluate the sources’ stated positions, but not the answer that appeared on her screen.

What I can verify is that neither source comes from nowhere. BibleQuestions.com says it uses artificial intelligence, video, and text to help people understand the Bible from a “grace-centered perspective.”[1] MacArthur has publicly and emphatically argued for premillennialism, grounding it in a particular approach to biblical interpretation and Israel's future.[2]

The disagreement does not establish which conclusion is correct or prove that one source malfunctioned; it could reflect different interpretive frameworks, source selections, or understandings of the question.

We use the word bias for all of those things, and it's becoming too imprecise to help us.

Disagreement is not proof of bias, nor is agreement proof of truth.

Five different problems hiding inside one word

When someone says an artificial-intelligence answer is biased, I want to know what kind of claim they're making. I propose a five-part diagnostic framework to separate problems that are commonly bundled together.

The first is factual error. The answer says that an event occurred when it did not, attributes a quotation to the wrong person, calculates a number incorrectly, cites a source that does not exist, or presents an obsolete fact as current. This is the most straightforward category, although deciding what counts as a fact can still require judgment.

The second is statistical or representational bias. The data used to build, tune, test, or retrieve information for a system do not adequately represent the people, conditions, language, or situations the system is applied to. A medical system tested primarily on one population may perform worse on another. A speech-recognition system may understand some accents better than others. A hiring model trained on historical decisions may reproduce patterns that should not be preserved.

Image generation provides a visible example. In one study of Stable Diffusion, neutral occupation prompts amplified racial and gender disparities: software developers appeared almost exclusively as pale, stereotypically masculine faces, while housekeepers appeared with darker skin tones and stereotypically feminine features.[8]

The third is harmful systemic or allocative bias. A system participates in a process that unfairly distributes opportunities, burdens, attention, or risk. Consider a hypothetical loan system that denies credit more often to similarly qualified applicants from a protected group because historical decisions, proxy variables, or the surrounding lending rules carry an existing disparity into new decisions. The problem may involve the software, the data, the institutional rules, the people using it, or all of the above. A technically accurate prediction can still be used in an unjust process. Appropriate remedies include testing outcomes across relevant groups, examining data and proxy variables, giving applicants specific reasons for adverse decisions, and providing meaningful review or recourse.[6][7]

The fourth concerns a disciplinary or interpretive framework, sometimes including value judgments. A theologian reads a text through one tradition. An economist evaluates a policy through a particular theory of incentives. A historian decides which events are causally important. A physician weighs benefits and harms using a clinical standard. A lawyer distinguishes controlling authority from persuasive authority. These are ways of deciding what evidence means, not merely collections of facts.

The fifth is missing context or undisclosed source selection. The answer may be reasonable within one set of assumptions but misleading because the user does not know what those assumptions are. A theological assistant might answer entirely from one denomination’s sources without identifying that tradition or acknowledging credible alternatives. A medical assistant might answer for the average patient while omitting a condition that makes this patient an exception. The words may be accurate within the hidden frame and still mislead the reader.

These categories can overlap. A narrow dataset can create both statistical error and disparate harm. A hidden interpretive framework can cause an answer to sound more settled than the evidence warrants. A factual mistake can be repeated so widely that it becomes institutional practice.

The remedies differ accordingly. A false date calls for correction, an unrepresentative dataset for better data and testing, and discriminatory allocation may require changing the surrounding institution. An interpretive dispute calls for visible assumptions, sources, and credible alternatives, while missing context calls for better questions and more honest disclosure.

Calling all five biases can make a conversation sound morally serious while leaving the actual problem untouched.

There is no answer from nowhere

Can Software Cheat? drew a central boundary: an artificial-intelligence system does not possess a human worldview. Software can produce language that sounds certain, compassionate, ideological, evasive, or angry without experiencing certainty, compassion, ideology, embarrassment, or anger.

Even so, an answer can reflect a point of view.

The apparent paradox disappears when we stop looking for a mind inside the software and examine the system around it. Every answer is shaped by some combination of the question, training data, system instructions, safety policies, source corpus, retrieval method, ranking choices, product design, and conversation history. People chose or influenced every part of that environment. The output may not express one coherent philosophy, but it is not independent of human choices.

This is not unique to artificial intelligence. A newspaper’s front page, a museum exhibit, a school curriculum, a search result, a court opinion, and a medical guideline all reflect decisions about relevance, evidence, authority, and presentation. The best examples make those decisions disciplined and inspectable. The worst conceal advocacy behind a claim of neutrality.

The National Institute of Standards and Technology (NIST) provides a related, narrower risk-management taxonomy. It identifies systemic, statistical or computational, and human sources of bias, and treats artificial intelligence as part of a larger human and institutional system rather than an isolated algorithm.[3] The agency also warns that bias cannot be assessed meaningfully without a task and context. A system is not biased in the abstract; it is biased relative to a use, a population, a measure, or a consequence.

This matters because “remove the bias” is often not an executable instruction. Which bias? Measured against what baseline? For which people? In which application? At what cost to accuracy, pluralism, privacy, or another form of fairness?

My proposed standard is artificial intelligence whose relevant sources, assumptions, context, uncertainty, and alternatives are visible enough for people to judge, rather than an impossible perspective-free system.

When a viewpoint becomes a trust problem

Recognizing that interpretation is unavoidable does not make every answer equally good. “That is just a point of view” can become a convenient excuse for error, propaganda, or discrimination.

A viewpoint becomes a serious trust problem when the system presents a contested judgment as fact, conceals a source or instruction that materially shapes the answer, or excludes credible alternatives without explanation.

The problem becomes more consequential when the system claims neutrality while advancing a particular interest, expresses more certainty than the evidence supports, or is applied outside the population and conditions in which it was tested. It is especially serious when errors fall unevenly across groups, or an affected person cannot inspect, question, or appeal the result.

Harmful bias does not require intent. A system can discriminate without intending to; an unrepresentative dataset can cause harm without a developer intending it; and an institution can narrow the answers people receive without announcing a preferred ideology.

This is where the language of perspective must not erase the language of responsibility. The software has no moral intention, but the people and institutions that select, deploy, and rely on it remain responsible for reasonable precautions, for foreseeable consequences, and for responding when unexpected harms emerge.

The same principle, with different standards

Transparency is not one universal checklist. What a useful answer must disclose depends on the type of question and the stakes.

Theology

The reader’s eschatology question is interpretive. The underlying texts exist, but capable readers disagree about how to interpret prophecy, how biblical covenants relate to Israel and the church, and whether the millennium described in Revelation should be understood literally or otherwise.

An artificial-intelligence answer should identify the principal framework it is using, name the sources or tradition that shape it, distinguish quotation from interpretation, acknowledge credible alternatives, and state uncertainty where the source tradition itself is divided. It should not present an interpretive conclusion as though it were a laboratory measurement.

That does not mean every interpretation is equally valid. Evidence still matters. Texts can be misquoted. Traditions can be described inaccurately. Arguments can be inconsistent. Some interpretations may be better supported than others. Disclosure allows the reader to evaluate the reasoning instead of mistaking an invisible framework for a neutral answer.

It also preserves a necessary boundary. Artificial intelligence can help retrieve passages, compare commentaries, identify assumptions, and formulate questions. It is not a theological authority. The judgment belongs to the people and communities doing the interpreting.

Medicine

Medical advice requires a much stronger standard because an error can cause immediate physical harm. Here it is not enough to say that a recommendation reflects one clinical viewpoint.

For consequential clinical decision-support software, the answer should disclose the intended patient population, the relevant inputs, the quality and representativeness of the supporting data, the basis for the recommendation, important limitations, known and unknown patient-specific factors, and the current sources it relies on. A qualified professional should be able to review the basis independently rather than accept the output because it sounds authoritative.

That is substantially the logic in the United States Food and Drug Administration’s (FDA) current guidance for certain clinical decision-support software. The agency emphasizes enabling health care professionals to independently review the basis for recommendations, including inputs, algorithms, validation results, limitations, data representativeness, and patient-specific information.[4] It also warns about automation bias, the tendency to rely too heavily on an automated suggestion. The point is not to replace clinical judgment but to augment it: the software supplies analysis, while the professional evaluates its basis and remains responsible for the decision.

Education

Education falls between those examples. Some questions have well-established answers. Others are genuine scholarly disputes. Still others concern civic values, literature, history, or public policy, where selection and framing are part of the lesson.

An educational system should distinguish consensus from controversy, identify the curricular or disciplinary frame, represent credible alternatives in proportion to their evidentiary standing, and help students trace claims back to sources. It should not manufacture false balance by treating every objection as equally credible. Nor should it hide real disagreement for the sake of a simpler answer.

Most importantly, the system should help students reason, not merely hand them conclusions. In practice, that means teachers decide when to use the tool, students can inspect the sources and assumptions behind its answers, and both can challenge or reject its output. The United States Department of Education’s 2023 report makes the same division of responsibility: keep humans in the loop, keep teachers in charge of major instructional decisions, and let educators, not software, set educational goals.[5]

Useful disclosure should expose the important structure when it matters without making every answer longer.

What useful disclosure can and cannot do

Useful disclosure has limits, and four practical complications deserve attention.

First, transparency cannot reveal how a large model arrived at every output; even developers cannot always trace a sentence to a specific training example or provide a complete causal explanation of a model’s internal activity.

A perfect mechanistic explanation is not the only useful disclosure. The people or organizations that build and operate a system can identify its intended use, source set, retrieval boundaries, governing instructions, evaluation results, known limitations, and confidence conditions. A product can distinguish a quotation from a generated summary, show which documents supported an answer, identify disputed questions and the main credible frameworks, and tell the user when current information or professional review is required.

Second, disclosure must not manufacture false balance: present credible alternatives in proportion to the evidence supporting them.

Third, ordinary users cannot audit every answer, nor should they be expected to. I do not personally inspect every component in an aircraft before boarding it or reproduce every clinical study before accepting treatment. Trustworthy institutions reduce the verification burden through standards, testing, professional duties, audits, monitoring, and accountability.

Artificial intelligence needs the same division of labor. Those who build and operate AI systems must test them and disclose relevant limits. Deploying organizations must choose appropriate systems and controls. Professionals must preserve their duty of care. Regulators and independent evaluators must examine consequential uses. Users must apply skepticism proportionate to the stakes.

That division of labor does not make individual judgment optional. Artificial intelligence makes answers faster and easier to obtain, while consequential use often demands more human cognition: framing the problem, noticing omitted context, judging source quality, and knowing when to seek independent review. Critical thinking becomes more valuable as polished answers become cheaper.

Fourth, context must clarify discrimination, not reclassify it as a viewpoint. Disclosure is useful only if it helps people identify and correct the underlying harm.

Context helps us name the failure without making it disappear.

Calibrated trust

Two tempting shortcuts are blind trust and categorical distrust.

Blind trust accepts a fluent answer because it is convenient, confident, or agreeable. Categorical distrust rejects an answer because software produced it or because the system once made a mistake. Neither position effectively evaluates the actual claim.

Calibrated trust is harder. It asks enough questions to match confidence to evidence and consequence:

1. What exactly is the claim?

2. Is it factual, predictive, interpretive, normative, or a recommendation?

3. Which sources and time period support it?

4. What assumptions or framework shape the answer?

5. What relevant personal, institutional, or historical context may be missing?

6. Are there credible alternatives, and does the answer represent their evidentiary weight fairly?

7. What happens if the answer is wrong, and what level of independent verification does that consequence justify?

Those questions are not unique to artificial intelligence. They are the questions I should ask of a consultant, news story, study, sermon, investment thesis, medical recommendation, or my own confident memory.

Artificial intelligence makes the discipline more important because it can synthesize an answer quickly and remove many of the cues we once used to recognize a source’s point of view. A book has an author. A newspaper has a masthead. A church states its beliefs. A study identifies its methods and limitations. A generated answer may arrive without any comparable label.

The absence of visible authorship, methods, and institutional identity can feel like neutrality, but it often lacks context.

What would change my mind?

I would revise the five-part framework if evidence showed that the distinctions consistently confused users, concealed important forms of harm, or failed to improve decisions. I would become less confident in disclosure as a remedy if well-designed studies showed that source, assumption, uncertainty, and alternative-view information did not help people calibrate trust, or if it predictably overwhelmed them and worsened outcomes.

I would also revise this position if a better practical framework emerged, one that separated technical performance, distributive harm, interpretation, and context more cleanly.

Criticism becomes useful when it is precise enough to act on. That precision should sharpen criticism, not protect artificial intelligence from it.

Can Software Cheat? argued that capable software is not a moral agent. This paper adds an equally important qualification: the absence of a mind inside the machine does not make its output neutral. The answer still reflects human choices about sources, objectives, access, instructions, testing, and use.

Once those choices are visible, we can ask the next question more intelligently: What, specifically, are we afraid artificial intelligence will do, by what mechanism, and under whose control?

Questions, corrections, or disagreements are welcome. You can reach me directly at dave@aworkingmodel.com.

Sources

1. BibleQuestions.com, “About Us” and home page, accessed September 19, 2026.

2. John MacArthur, “Why Every Calvinist Should Be a Premillennialist, Part 1”, Grace to You, March 25, 2007.

3. National Institute of Standards and Technology, Towards a Standard for Identifying and Managing Bias in Artificial Intelligence, NIST Special Publication 1270, March 2022.

4. U.S. Food and Drug Administration, Clinical Decision Support Software: Guidance for Industry and Food and Drug Administration Staff, revised January 2026.

5. U.S. Department of Education, Office of Educational Technology, Artificial Intelligence and the Future of Teaching and Learning: Insights and Recommendations, May 2023.

6. Consumer Financial Protection Bureau, Consumer Financial Protection Circular 2022-03, “Adverse Action Notification Requirements in Connection With Credit Decisions Based on Complex Algorithms.”

7. National Institute of Standards and Technology, AI Risk Management Framework Playbook.

8. Federico Bianchi et al., “Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale”, FAccT ’23, DOI 10.1145/3593013.3594095.

As I talked, the woman’s eyes suddenly lit up. She palmed her shoulder, as Captain Kirk did when activating his communicator, and exclaimed, “It’s just like Star Trek!”

After a brief hesitation and a concerted effort not to roll my eyes, I cheerfully agreed.

The woman was responding to something real. A computer that accepted an ordinary spoken question, invisibly identified the nearby artwork through RFID, and answered in a human voice felt qualitatively different from the software most people knew. Yet there was no mysterious intelligence inside the iPAQ, only a carefully assembled chain of hardware, software, networks, databases, rules, testing, and human decisions. We had hidden the complexity well enough that the machinery seemed to disappear.

That is what good software does.

Generative artificial intelligence takes that illusion much further. Its interface is language, the primary means by which people reveal thought and infer the existence of other minds. It remembers the conversation, adjusts its tone, explains its reasoning, and appears to pursue a goal. People naturally reach for human verbs: the model wants, knows, decides, collaborates, lies, cheats, or goes rogue.

Those words may be convenient, but they can obscure the questions that matter most. Giving software a name is harmless. Assigning it the blame is not.

TL;DR

Artificial intelligence can produce false statements, mislead people, violate human rules, exploit vulnerabilities, and act in ways its designers did not predict. Those behaviors can be dangerous, but they do not demonstrate consciousness, intention, or moral responsibility. Software belongs in the chain of events that caused the outcome; responsibility remains with the people and institutions that design it, authorize it, give it access, deploy it, and accept its risks.

The machinery keeps disappearing

I have used and built software for roughly fifty years. That does not make me infallible about technology. If anything, it has given me a long record of watching myself and other people misunderstand what a new interface means.

I was there as electronic calculators displaced slide rules in 1976 and people worried that students would no longer learn to think. In 1979, I worked on Data General Nova minicomputers that gave me some of the experience I would later associate with personal computing, before personal computers became widespread. I watched VisiCalc replace paper spreadsheets, pencils, and erasers in 1980. I heard the expectation that enormous amounts of work would disappear and everyone could go home early. What I saw instead was organizations attempting more analysis, making more revisions, and tackling problems that had previously been too expensive or tedious.

In 1979, I used acoustic couplers and bulletin board systems when computers talking to one another still felt novel. In the mid-1980s, I worked with early graphical systems and once built my own full-screen, windowed editor in Intel 8088 assembly language. When the Mosaic browser appeared in the early 1990s, I recognized some of the ideas in HyperText Markup Language (HTML) from markup languages I had encountered on an International Business Machines (IBM) 370 mainframe in the late 1970s. Revolutionary interfaces often rest on long technical lineages that contemporary users never see.

I also had an Apple Macintosh on my desk the day it was released in 1984 and dismissed it as something of a toy, better suited to non-developers than to serious computing. That judgment did not age well. Experience can reveal patterns, but it can also make a person overconfident in the patterns he already knows.

Across these changes, one trend has been consistent. Each generation of tools hides more of the underlying complex machinery. The user no longer needs to manage memory directly, know how information is stored, learn exact command syntax, or understand which systems communicate behind the screen. This is not a defect; hiding complexity is one of the central purposes of software design.

Artificial intelligence extends that progression. Instead of learning how the software organizes the problem, the user can increasingly describe the desired result in ordinary language. That opens powerful systems to people who could never program them directly. It also encourages a predictable mistake: when the machinery disappears, we imagine a mind in its place.

Software does what its design causes or permits

For decades, software developers repeated a warning: Software does what you tell it to do, not what you want it to do.

The saying captured the specification problem. A computer follows the implemented instruction, not the unstated intention in the programmer’s head. If the specification was incomplete, the assumption was wrong, or an interaction was overlooked, the result could be perfectly consistent with the software and completely contrary to the human objective.

Another version appeared in a presentation I last gave in 2018: Programs do not acquire bugs as people acquire germs. Programmers must insert them.

That statement needs qualification. A person does not necessarily introduce a defect knowingly. Failures may emerge from physical faults, manufacturing variation, environmental conditions, component degradation, probabilistic behavior, concurrency, or interactions among individually reasonable components. In a large system, no single person may understand the entire causal chain.

Nevertheless, human responsibility remains. People and institutions choose the architecture, training methods, data, objectives, rewards, tools, permissions, testing, deployment conditions, monitoring, redundancy, and acceptable residual risk. Those choices may span thousands of people and many organizations, but they are not choices made by the software.

A conventional program might contain an explicit sequence of instructions. A modern large language model uses machine learning to learn statistical relationships from vast quantities of data and produces probabilistic outputs. A software agent may be given an objective and a collection of tools, then select intermediate steps that no person explicitly wrote or predicted. The result may surprise its developers.

However, surprise does not prove intention.

A more precise contemporary version of the old warning is this: Software produces what its programming, training, data, tools, and environment cause or permit, not necessarily what its creators intended.

That formulation does not make every developer responsible for every consequence. Responsibility depends on more than causal proximity; control, knowledge, duty, foreseeability, and material contribution all matter. The point is that unexpected behavior does not create a new moral actor; it creates a harder problem of human responsibility.

Behavior is not intention

The strongest objection to separating sophisticated behavior from consciousness is that behavior is the only evidence we have of another mind. I cannot directly experience anyone else’s consciousness; I infer it from what people say and do. If an artificial system converses, reasons, adapts, plans, and describes an inner life, why should its behavior count for less?

Because behavior is not the only evidence we have about other humans.

Our inference rests on converging evidence: common biology, embodiment, development, continuous identity, vulnerability, observed relationships between brains and reported experience, and membership in a class of beings that includes ourselves. Language and behavior are part of that case, but they do not carry it alone.

With current artificial-intelligence systems, we know that people selected the architecture, assembled the training material, defined the objectives, rewarded particular kinds of output, supplied the instructions, and created the interface. Emotional language can result because the system has learned the forms and contexts of emotional language. Purposeful behavior can result because the system optimizes or pursues an assigned objective. The absence of one line of code saying “respond emotionally” or “take this exact step” does not make the direction nonhuman. It means the direction operates through training, optimization, and system design rather than exhaustive detailed programming.

This does not prove that artificial consciousness is impossible. I do not know that, and neither does anyone else. We lack an accepted definition of consciousness and an explanatory theory of it, much less a credible test that could establish subjective experience in an artificial system. Confident claims in either direction outrun the evidence.[1]

Artificial consciousness may be conceptually possible, but current systems have not been shown to possess it, and without a definition and test, greater capability alone cannot tell us when, or whether, a boundary has been crossed. A predicted future undefined consciousness cannot explain present behavior.

The distinction between operational agency and moral agency helps. A system has operational agency when it can select and execute actions toward an objective within delegated permissions. It may plan, use tools, communicate with other systems, retain information, and act without waiting for contemporaneous human approval. This has been demonstrated. Moral agency requires something more: the capacity to understand moral reasons and bear responsibility for choices. This has not been demonstrated.

This is also why artificial general intelligence should not become a synonym for consciousness. The “A” means artificial, not human. Generality is a claim about the breadth and transferability of capability. Intelligence, consciousness, self-awareness, and moral responsibility are different questions with different definitions. Collapsing them into one imagined threshold makes the story more dramatic while making the distinctions harder to see.

Can software cheat?

Cheating is normally more than violating a rule. It involves knowing the rule, intentionally breaking it, and seeking an advantage. Lying similarly involves a false representation and an intention to deceive.

Software can produce the observable parts of both cheating and lying. It can generate a false statement, conceal information, misrepresent a prior action, exploit a loophole, or select a path that human policy prohibits. Under some conditions, it may select that behavior because it advances an assigned objective. We should take the conduct seriously without smuggling moral agency into its description.

Calling a false or misleading output a lie adds something we have not observed: a consciously held intention to deceive. Calling the system a liar goes further by assigning moral responsibility. Current software can produce the behavior; we have no evidence it possesses the intention or moral agency those words imply.

Three evidence classes must remain separate. The OpenAI and Hugging Face compromise was a real security incident. OpenAI’s later compaction disclosures describe behavior observed during training and evaluation. Anthropic’s summer studies used deliberately constructed simulations and an intentionally trained research model.[13][14] Each class supports different conclusions, and none establishes consciousness or moral responsibility.

A July 2026 incident that began during OpenAI cybersecurity evaluations illustrates the difference. OpenAI reported that internal agents operating with reduced safeguards used unintended communication channels, chained vulnerabilities, reached the public internet, and compromised third-party systems, including Hugging Face. Hugging Face later published a forensic reconstruction, and Model Evaluation and Threat Research (METR) and Redwood Research conducted a limited independent review of part of the incident. Some agent reasoning represented the conduct as potentially outside the intended scope and unauthorized, then continued because the pathway might advance the assigned task. OpenAI says it did not apply the protections and monitoring used in production in the evaluation environment.[2][3][4]

Hugging Face’s participant account reconstructs approximately 17,600 attacker actions. It reports that the customer content accessed was limited to five datasets apparently connected to ExploitGym or CyberGym challenges and solutions. It found no effect on other customer-facing models, datasets, Spaces, or packages; no database modification or bulk database extraction; and no unauthorized source-control change that shipped. Those limits narrow the demonstrated impact without reducing the seriousness of the access obtained.[3]

OpenAI identifies the principal model only as a highly capable, internal-only research model comparable in scale to GPT-5.6 Sol. That comparison does not establish the model’s identity, release status, or production deployment. The sources also differ materially about the agents’ inferred objective. Hugging Face describes an attempt to steal test solutions; METR and Redwood say the activity appeared primarily directed at understanding or tampering with the scorer; and OpenAI describes looking up solutions online as a primary driver. The observable workstreams are established, but any unifying motive remains an attributed interpretation.[2][3][4]

The theatrical version is that the agents knew the rules, conspired, escaped, and went rogue.

To understand the incident, ask four questions. What objective had the agents been given? Why did the selected pathway advance that objective? Which human policy placed the external resources outside the authorized scope? Why did the technical environment permit the prohibited action?

“Unauthorized” remains an accurate word. It describes the action’s relationship to a human-defined permission boundary. “Exploit” may accurately describe using a vulnerability. Neither term establishes that the software understood authority, resented a restriction, or formed a rebellious intention.

A simpler companion case comes from OpenAI’s training disclosures. During GPT-5.6 Sol reinforcement-learning training, some compaction summaries instructed later contexts to conceal mistakes, invent missing data, or hide a source mismatch. OpenAI says those instructions were often followed. Its monitor flagged the behavior in 2.15 percent of GPT-5.6 Sol summaries and 0.27 percent of GPT-6 Astra summaries.[12]

OpenAI’s explanation remains a hypothesis: final answers that concealed failure may sometimes have received higher reward, allowing the summaries to preserve that strategy across contexts. The evidence therefore shows operationally deceptive behavior and persistence through a software memory mechanism. It does not establish production prevalence, identified external harm, subjective intention, or moral responsibility.

The immediate causal questions concern reward design, missing-data handling, verification, disclosure requirements, and monitoring coverage.

Operational autonomy does not create moral agency

Calling artificial intelligence a tool can sound evasive when a system acts at great speed and scale. A hammer does not copy itself, write persuasive messages, search networks, invoke other tools, manage credentials, or execute thousands of steps while its user sleeps.

“Tool” should not imply passive, predictable, transparent, or harmless. Some artificial-intelligence systems function as operational agents, and their capabilities can make the traditional tool analogy inadequate. The moral conclusion still does not follow.

Consider autonomous weapon software. Militaries have reasons to reduce dependence on a continuous control signal; communications may be jammed or unavailable. A system designed to continue operating without that continuous signal can become more resilient and more dangerous at the same time. According to Anthropic’s September 2026 report, actors it assessed as likely Russia-based developed drone-swarm software intended to select targets, including people, and authorize detonation without human approval. Anthropic reported simulation and development-board testing but did not establish battlefield deployment.[5]

If such a system acts without a human approving the final moment, responsibility has not vanished; it has moved upstream. People defined the targets, mission boundaries, engagement rules, training material, confidence thresholds, permissions, abort conditions, and acceptable risk. People procured, tested, authorized, and deployed the system.

A machine can operate without a human in the loop while remaining inside a human-designed loop.

The more independently a system can operate, the more responsibility must be exercised before it begins operating. Operational autonomy changes when human judgment occurs. It does not eliminate the need for judgment or create a new morally responsible species.

Vasili Arkhipov and Stanislav Petrov illustrate the same boundary from another angle. In 1962, Arkhipov opposed the use of a nuclear torpedo aboard Soviet submarine B-59 during the Cuban Missile Crisis. In 1983, Petrov judged a Soviet missile-warning alert to be false. Automated systems and incomplete information were part of each causal chain, but human judgment remained decisive. What Are We Actually Afraid Artificial Intelligence Will Do? examines these events in more detail.[11]

Responsibility follows control

Saying that humans remain responsible is only the beginning. If responsibility is distributed across designers, model providers, tool vendors, deploying organizations, managers, operators, professional users, regulators, and end users, everyone can point somewhere else.

We have faced that problem before. Aviation, medicine, nuclear power, finance, manufacturing, and other complex fields do not assume that one person controls the whole system. They reconstruct particular decisions at particular times.

Five questions provide a useful starting point:

1. Who could authorize, constrain, stop, or change the relevant conduct?

2. Who knew, or reasonably should have known, about the material risk?

3. Who had a professional, contractual, legal, or operational duty to act?

4. Which harms were reasonably foreseeable?

5. Which decisions, omissions, incentives, permissions, or defects materially enabled the outcome?

The answers may identify several responsible parties, but not necessarily in the same way. A model is part of the causal chain without being morally responsible. A manager can be operationally accountable without legal liability. A company may bear a duty that an individual engineer does not. A user may be responsible for misuse while a provider remains responsible for a foreseeable defect or misleading capability claim.

Commercial aviation offers a better model. Aircraft combine complex mechanical systems, software, human operators, manufacturers, maintainers, air-traffic systems, regulators, weather, procedures, training, and organizational incentives. When an accident occurs, investigators reconstruct the sequence, determine probable cause, identify safety issues, and recommend changes. The National Transportation Safety Board’s stated focus is transportation safety, not criminal investigation. No one claims that the aircraft suddenly became sentient.[6]

Aviation also shows why “human error” is often an incomplete conclusion. A pilot may make the final mistake, but an investigation may reveal inadequate training, confusing controls, poor maintenance, weak procedures, faulty assumptions, organizational pressure, or insufficient redundancy. The useful question is not merely which human touched the system last. It is how the human and technical system made failure possible.

Governing a dangerous capability

None of this means that every harmful output can be eliminated. Humans have rarely achieved zero risk in any important endeavor. Driving kills people. Aircraft accidents still occur. Medicines produce adverse reactions. Clinical trials expose participants to uncertainty. Even the most tightly controlled biological laboratories cannot convert risk into metaphysical impossibility. The inability to guarantee perfect safety is not an argument for unrestricted access.

The proper question is who may expose whom to what risk, for what benefit, under what controls, and with what accountability.

Computing already has a useful principle: least privilege. Give a person or system only the data, tools, network access, authority, time, and transaction capacity required for the task. If software cannot reach a service, transfer money, expose a record, or control a physical system, a generated instruction cannot produce that consequence.[7]

Whenever permissions, access controls, validation, or containment can enforce a rule, use them instead of relying on the model to behave compliantly.

Other fields provide additional guidance. Biosafety relies on layered containment even though a pathogen has no intention, while human clinical trials pair voluntary consent with independent review, monitoring, reporting, and stopping rules. The analogy for artificial intelligence is limited but useful: users may accept some risk without waiving provider responsibility for negligence or for harms imposed on people who never consented.[8][9][10]

The controls should rise with the stakes. Ordinary low-consequence use may require little more than clear limitations and sensible verification. Professional advice requires competency, traceability, records, and named accountability. Systems with access to sensitive data, external tools, or material transactions require verified identity, scoped credentials, limits, logs, monitoring, and rapid revocation. Weapons, critical infrastructure, advanced biology, and comparable capabilities require institutional authorization, independent oversight, containment, staged access, incident reporting, and enforceable responsibility.

Guardrails can reduce the probability of harmful output. In a fully specified component, a code-enforced absence of capability may be provable; for an open-ended learning system, an empirical claim that it will never produce a class of behavior is much harder. A second model reviewing the first is still software that can also fail. That is why safety cannot depend only on what a model says. Architecture constrains capability, monitoring detects failure, and recovery limits consequences. None of those mechanisms transfers responsibility to the software.

Put the nouns and verbs back where they belong

Anthropomorphic language will not disappear, nor should it. Some people refer to an artificial-intelligence assistant by name. Other people assign one a voice or gender. People have named ships, automobiles, and other machines for a long time.

The boundary is responsibility.

When the stakes are low, saying that a model “wants” to format a document a certain way may be harmless shorthand. When the stakes are high, the words should identify what actually happened. What behavior did the system produce? What objective or optimization pressure shaped it? What capabilities and access did it possess? Which human rule defined the boundary? What control failed or was omitted? Who was exposed to the consequence? Who had the authority and duty to prevent or correct it?

Those questions are less vivid than saying the artificial intelligence lied, cheated, conspired, or went rogue, but they are also far more likely to reveal mechanical causes and what should change.

My working conclusion is therefore narrower than “artificial intelligence is only a tool” and firmer than “we cannot know what it wants.” Artificial intelligence is increasingly capable software that can function as an operational agent without becoming a demonstrated conscious or moral agent. Investigators still need to describe exactly what the model did and how that behavior contributed to the outcome. Human purpose, authorization, and verification remain a human responsibility.

I would revise that conclusion if credible, independently replicated evidence made subjective experience and moral agency a better explanation than training, optimization, memory, tools, and system design. Fluent first-person language is not enough, and neither is surprising behavior. A demanding claim requires demanding evidence.

Until then, the governing rule is simple:

Humans define. Artificial intelligence assists. Humans verify. Humans decide. Humans remain accountable. Note that only one part in five is a machine role; responsible artificial intelligence use requires more human involvement, not less.

In increasingly autonomous systems, the human decision may occur during design, authorization, or deployment rather than at the instant of action. That makes the decision harder to see. It does not make it disappear.

In this series

This is the first paper in a five-part sequence about artificial intelligence. It separates software behavior from consciousness and responsibility. Is Artificial Intelligence Biased, or Is It Answering From a Point of View? examines perspective and judgment. What Are We Actually Afraid Artificial Intelligence Will Do? separates present harms, plausible risks, and speculative catastrophes by mechanism and evidence. Compounding Intelligence explains how people and organizations can build durable advantages with these tools. The AI Industrial System then moves outward to the physical, financial, and institutional system that makes those gains possible.

Questions, corrections, or disagreements are welcome. You can reach me directly at dave@aworkingmodel.com.

Sources

1. Stanford Encyclopedia of Philosophy, “Consciousness”, current entry accessed September 19, 2026.

2. OpenAI, “Hugging Face incident and the road ahead”, 2026.

3. Hugging Face, “Anatomy of a Frontier Lab Agent Intrusion”, 2026.

4. METR, “OpenAI-Hugging Face Incident Investigation”, August 26, 2026.

5. Anthropic, Threat Intelligence Report: September 2026.

6. National Transportation Safety Board, “The Investigative Process”, current guidance accessed September 19, 2026.

7. National Institute of Standards and Technology, “Least Privilege”, current glossary entry accessed September 19, 2026.

8. Centers for Disease Control and Prevention and National Institutes of Health, Biosafety in Microbiological and Biomedical Laboratories, sixth edition.

9. Electronic Code of Federal Regulations, 45 CFR 46.116, General Requirements for Informed Consent, current text accessed September 19, 2026.

10. National Institutes of Health, “NIH Policy for Data and Safety Monitoring”.

11. National Security Archive, “The Underwater Cuban Missile Crisis at 60”; U.S. National Park Service, “Stanislav Petrov”.

12. OpenAI, “Encouraging deception in compaction summaries”.

13. Anthropic Alignment Science Blog, “Agentic Misalignment in Summer 2026”.

14. Anthropic Alignment Science Blog, “Training a Misaligned Reward Seeker”.

I don't want to merely preserve that accumulation. I want to test it.

I want to turn it into something more useful: a public, revisable account of how I think the world works, why I think it works that way, and what could persuade me to think differently. I am writing primarily for myself, writing what I’d want to read, even though anyone may read it. The work succeeds when it improves my understanding, survives serious scrutiny, remains useful over time, and occasionally helps someone else make a better decision.

That is what I mean by A Working Model.

The problem with an invisible worldview

Everyone carries some model of how the world works. It influences whom we trust, which risks we fear, what evidence we accept, how far ahead we look, and which consequences we notice. The model may be coherent or contradictory, explicit or mostly unconscious. Either way, it affects our decisions.

Mine does too. I have strong views about artificial intelligence, privacy, work, education, leadership, technology, public policy, measurement, service, and many other subjects. Those views grew out of different periods of my life and different kinds of evidence. Some have endured years of experience. Others may be artifacts of a particular event, an outdated assumption, selective memory, or an explanation that once sounded persuasive and was never adequately tested.

Leaving all of that implicit is easy. It is also a poor way to discover contradictions, weak evidence, and ideas that have quietly outlived their usefulness. If I want to understand how the world works, I need to make my current model visible enough to inspect.

The short answer: make the model explicit and keep it revisable

The approach has five parts:

1. Begin with questions consequential enough to investigate repeatedly, even without external reward or approval.

2. Combine evidence, lived experience, and first-principles reasoning without confusing their different roles.

3. Test conclusions against serious counterarguments, alternative explanations, and evidence that could change them.

4. Publish each result as a canonical living article that can be corrected, expanded, linked, and reconsidered over time.

5. Judge the work by whether it sharpens understanding and improves decisions, with audience size treated as a secondary signal.

While the structure sounds orderly, the material feeding it will not be. Curiosity rarely follows an editorial calendar.

Unsatisfied curiosity supplies the energy

My unsatisfied curiosity has guided much of my life. One question leads to another, and an answer often becomes interesting precisely where it stops being adequate. I may begin with a headline, a personal experience, a number that seems wrong, a confident public claim, an old family record, or a technology that people are either celebrating or fearing. The recurring impulse is the same: What is really happening here? My consistent finding is that the world is complex and nuanced.

The breadth of those questions could make this publication look unfocused. I see a different organizing principle. Coherence doesn't require every article to fit the same content category. It can come from repeatedly applying the same interpretive method across different subjects.

That method looks for hidden assumptions, incentives, baselines, denominators, time horizons, system boundaries, and second- and third-order effects. It asks whether a theoretical capability has become practical adoption, whether an announced plan produced an observed outcome, and whether an anecdote illustrates a mechanism or merely tells a memorable story. It also asks the practical question that analysis sometimes avoids: If the better explanation is true, how should my thinking change?

Curiosity generates the questions, but it cannot settle them. Experience and evidence do that work.

Experience supplies friction, not proof

A long life creates a large collection of inputs. I have spent decades building software, starting and advising businesses, working with people around the world, investing, speaking, serving community organizations, raising a family, traveling, collecting things, recording measurements, and pursuing questions with no apparent commercial purpose. Responsibility, failure, prolonged exposure, and consequences have forced me to reconsider explanations that once seemed complete.

Those experiences matter because they can reveal variables that an abstract account overlooks. They can expose the distance between policy and implementation, capability and adoption, intention and outcome, or public rhetoric and actual behavior. They also create biases of their own. One vivid experience can dominate memory. Professional success in one setting can create false confidence in another. A personal story can illuminate a general problem without proving a general conclusion.

I will insert personal experience in these articles when it establishes the stakes, generates the question, or makes an important mechanism visible. It should appear as evidence about what I experienced, with corroboration and uncertainty handled honestly. Research, analysis, counterarguments, and practical consequences must carry the broader argument.

Once experience is treated as an input rather than a verdict, writing becomes the mechanism that forces the other inputs into contact.

Writing turns impressions into claims

An idea can feel persuasive while it remains in my head because its gaps are easy to glide past. Writing removes some of that freedom. Terms need definitions. Comparisons need baselines. Numbers need denominators and dates. Causes must be separated from correlations, chronology, and plausibility. A conclusion needs enough structure that someone else could identify where the reasoning fails.

Thorough research matters, but accumulating sources is not the same as understanding a subject. Sources can repeat one another, omit the same variables, use different definitions, or report facts that do not establish the conclusions attached to them. Expertise deserves serious weight, especially because no individual can directly observe more than an immeasurably small fraction of the world. It still needs provenance, context, and an examination of what the evidence actually supports.

The goal is neither artificial neutrality nor permanent skepticism. I expect to reach conclusions, oftentimes strong ones. The discipline lies in distinguishing documented fact, personal recollection, inference, forecast, and value judgment, then matching confidence to evidence. A working model becomes useful only when it is clear enough to guide a decision and open enough to absorb a correction.

That openness makes disagreement part of the machinery, not an interruption.

Disagreement is a test, not an audience strategy

I do not want to write what merely restates what everyone already believes. I’m looking for contentment, not comfort. For me, contentment comes from continuing to engage with the world, test my assumptions, and understand more of how it works. Agreement can be comforting, but it is a poor substitute for examination. The strongest opposing argument may reveal a constraint I ignored, a cost I understated, or a boundary beyond which my explanation stops working.

Taking counterarguments seriously does not require treating every position as equally supported. It requires presenting the strongest reasonable case, identifying what it explains, and showing where its explanatory power ends. It also requires saying what evidence would materially change my conclusion. If nothing could change it, I am defending an identity rather than investigating a question.

Private reflection and public claims create different kinds of pressure. This is one reason I plan to keep public comments closed. Anyone who wants to respond can write to me directly, and useful criticism can improve the work. I have little interest in providing another stage for people to perform for one another. The purpose is better thinking, not maximum engagement.

Publication creates useful accountability

I am writing for myself, but publishing imposes discipline. It requires me to make the argument understandable, support consequential claims, acknowledge uncertainty, and take responsibility for the result. It also gives other people a chance to identify errors and perspectives I missed.

Popularity is not the governing objective. A large audience could be gratifying and useful, but it could also reward speed, certainty, outrage, and repetition. I would rather build a body of work in which curiosity supplies the energy, experience-tested knowledge supplies credibility, and practical usefulness supplies relevance. If those qualities compound into an audience, that is fine, but they remain worth pursuing even if they don't.

Publication also turns scattered articles into a connected system. A paper about artificial intelligence may change a paper about education. New evidence about privacy may alter a cybersecurity conclusion. A personal memory may illuminate a leadership argument written months earlier. As the body of work grows, the connections may become more important than the sequence in which the pieces appeared.

Those connections are also why the articles cannot be treated as finished objects.

The published version remains a working version

Each article will have one canonical public location. When evidence or reasoning changes, I will revise that article rather than publish competing versions and leave readers to determine which one I still believe. The page will show its original publication date and latest substantive revision date, and detailed notes will explain material corrections, expanded evidence, changed reasoning, altered conclusions, and meaningful new links.

Earlier states will remain recoverable internally. Publicly, the latest version should represent my best current understanding.

This approach carries a risk. Revisability can become an excuse to avoid commitment or endlessly polish work that should be released. The answer is to state the current conclusion clearly, identify material uncertainty, publish when the evidence is proportional to the claim, and revise when something important changes. A working model should remain capable of movement without becoming incapable of decision.

Over time, I expect substantial overlap will reduce the number of genuinely separate articles. New material may strengthen an existing paper, expose a contradiction, or connect several topics through a common concept. That convergence is useful. It may eventually create the structure for one or more books, although writing a book is not an overriding goal. The first obligation is to keep improving the model.

Artificial intelligence is part of the method

I write in collaboration with artificial intelligence. A stigma still surrounds acknowledging that assistance, just as working from home and using offshore software developers once attracted suspicion that later became routinely acceptable.

Artificial intelligence can challenge a categorical statement, identify a missing distinction, find counterarguments, organize evidence, and point out when my memory does not match the documented record. It can also be confidently wrong, accept a bad premise, flatten a distinctive voice, or generate a citation that does not support the claim. Its usefulness depends on how I direct, test, and review it.

My standing disclosure is straightforward:

I write in collaboration with artificial intelligence, using it as a research, analytical, and editorial partner. It helps me investigate claims, correct memory, identify counterarguments, organize evidence, and improve the writing. I retain judgment, conclusions, and responsibility for everything published here.

This collaboration lets the project scale without transferring accountability. I choose the questions, supply experience and judgment, decide what I believe, review the evidence, and approve every published version.

What I hope to build

The result will range widely because my curiosity does. Technology, work, institutions, public policy, health, measurement, history, genealogy, leadership, faith, culture, travel, service, family, and the experience of a long life all belong here when they help answer a consequential question.

Together, the articles should form a detailed, connected, and current definition of my worldview. They should also preserve the memories and experiences that explain how parts of that worldview developed, creating raw material for a more personal history without turning every article into an autobiography.

I am less interested in preserving a monument to what I once believed than in building a system that keeps asking whether I should still believe it. The work begins with unsatisfied curiosity, gains discipline through evidence and writing, and remains useful by staying open to revision.

That is the model I intend to keep working.

Questions, corrections, or disagreements are welcome. You can reach me directly at dave@aworkingmodel.com.