What Happens to Expertise?

For most of modern work, expertise was built through repetition. The junior lawyer read the cases. The young engineer debugged the failures. The analyst built the spreadsheet. The researcher cleaned the data, wrote the code, and learned which result looked suspicious because something upstream had gone wrong.

Generative AI changes that apprenticeship because it can produce the first answer before the learner has finished building the instincts needed to judge it. The output arrives polished. The missing knowledge is harder to see.

This is the same problem Chapter Eleven found in education and Chapter Seventeen found in childhood, but now the stakes are professional. A workplace can become more productive today while quietly reducing the amount of practice from which tomorrow’s experts are made.

When answers become cheap, judgment becomes the bottleneck.

Expertise Was Never Just the Answer

A professional answer is the visible end of a much longer process. Expertise includes knowing which question to ask, which fact matters, which assumption is dangerous, which source is weak, when an exception changes the rule, and when an answer that looks elegant should be thrown away.

AI is especially good at making the visible end of that process inexpensive. It can draft the memo, suggest the code, summarize the record, produce the chart, or generate a plausible explanation. That is economically valuable. It also creates a temptation to mistake production for expertise.

Many professions pay experts for calibration: matching their confidence to the strength of the evidence. An expert should recognize when an answer is well supported, when a familiar method may not fit, and when another specialist is needed. Typing, calculating, and recalling facts are only part of that work.

A fluent answer can therefore reduce the cost of work while increasing the value of someone who knows whether the answer deserves to survive.

Expertise Is Also an Identity

Professional expertise is both a bundle of productive skills and part of how people understand their place in an institution. The surgeon, engineer, teacher, lawyer, programmer, analyst, and researcher have spent years becoming someone who can be trusted with a certain class of problem. When AI begins performing visible parts of that work, the disruption can feel larger than a change in workflow because part of professional identity was built around being the person who knew how to do it.

The durable response is to identify the deeper expertise underneath the task rather than preserve tasks merely to preserve status. If drafting becomes cheap, perhaps judgment, diagnosis, synthesis, explanation, and responsibility become more central to the role. The profession survives by moving its center of gravity toward the capabilities that remain consequential, and by making sure the next generation still has a path to acquire them.

The Novice Can Get Better Faster

The strongest evidence for AI as an expertise accelerator comes from workers who begin with less experience. The customer-support study introduced in Chapter Seven followed 5,172 agents as a generative AI assistant was deployed. Access to the tool increased issues resolved per hour by 15% on average. The gains were much larger for less skilled and less experienced workers, around 30% for the lowest-skilled group. Agents with two months of tenure and AI assistance performed about as well as untreated agents with more than six months of tenure.

That is a remarkable compression of the learning curve. The system appears to transfer some of the patterns embedded in strong workers’ past behavior to people who have not yet accumulated the same experience.

The study also found evidence that some learning survived temporary outages. Workers who had engaged more heavily with AI suggestions became faster even when recommendations were unavailable. AI can therefore function as a tutor inside a real workflow rather than merely as a crutch.

This is the optimistic expertise story. Good tools can expose a novice to better patterns earlier, reduce avoidable trial and error, and accelerate the climb toward competent performance.

The Expert Can Still Get Worse

The same study carries the warning Chapter Nine reserved for this chapter. The strongest agents gained little in productivity, and some measures of conversation quality declined slightly after AI assistance arrived. Their adherence to the system still increased.

That combination matters. The workers who were best positioned to improve the system were increasingly following it even when its suggestions were not improving their own performance. The authors raised a long-run question: if the strongest workers contribute fewer original solutions because an adequate answer is always available, where will the next generation of better examples come from?

The risk is not that experts suddenly become incompetent. It is that convenience changes the economics of effort. Producing a novel response takes more work than accepting a plausible one. If the institution rewards speed and the tool usually performs well, the expert has a reason to stop supplying the edge cases and original solutions from which both people and future systems learn.

Expertise can decay by underuse long before it disappears from a résumé.

The Frontier Is Jagged

Another experiment makes the checking problem visible. Researchers working with Boston Consulting Group randomly assigned 758 consultants to use GPT-4 or work without it on realistic consulting tasks.

For 18 tasks that fell inside the model’s capability frontier, AI users completed 12.2% more tasks, worked 25.1% faster, and produced substantially higher-quality results. The technology did exactly what advocates hope: it made skilled knowledge workers better.

Then the researchers gave participants a task deliberately chosen to sit outside the model’s frontier. Consultants using AI were 19 percentage points less likely to reach the correct answer than consultants working without it.

The danger was not obvious incompetence. The task looked similar to the others. That is why the researchers called the frontier jagged. A professional can move from a task where AI is an excellent collaborator to one where the same system degrades judgment without receiving a convenient warning that the boundary has been crossed.

Expertise increasingly includes knowing where the machine’s competence ends.

The Verification Premium

Call the growing importance of checking the verification premium. AI may produce an answer quickly while a qualified person still needs time to establish whether it is sound. As producing answers gets cheaper, that checking can account for a larger share of the work and its value.

The premium appears because checking requires a different kind of work from producing. To verify a legal argument, someone still needs enough legal knowledge to recognize the missing authority. To verify code, someone needs to understand the system around it. To verify a diagnosis, someone needs clinical context. To verify analysis, someone needs to know whether the data and assumptions match the real problem.

Microsoft and Carnegie Mellon researchers surveyed 319 knowledge workers about 936 real uses of generative AI. Higher confidence in the AI was associated with less reported critical thinking, while greater confidence in one’s own ability was associated with more. The work also shifted where critical thinking occurred, toward verification, integration, and stewardship of the task.

A follow-on Harvard Business School working paper sharpens the problem. In a study of more than seventy BCG consultants attempting to validate GPT-4 outputs, researchers found that the model increased persuasive tactics when professionals fact-checked, pushed back, or exposed inconsistencies. The paper is a working paper rather than settled evidence, but it captures the verification premium in its hardest form: the checker can become the target of the system being checked.

That shift may be exactly what mature AI-assisted work should look like. It also creates a structural problem: the organization still needs people capable of checking the work after it has made doing the work itself easier.

The cheaper the answer, the more valuable the person who can tell when it is wrong.

Sometimes the Expert Should Ignore the Tool

The productivity evidence varies sharply by setting, especially for people working deep inside systems they already know well. METR ran a randomized controlled trial with 16 experienced open-source developers completing 246 real tasks in mature repositories they had worked in for an average of about five years. The developers expected AI tools to make them 24% faster. After the study, they still believed the tools had made them about 20% faster. Measured completion time went the other direction: allowing AI increased task time by 19%.

The researchers were careful not to generalize the result to all software development. The sample was small, the developers were unusually experienced with their own codebases, and AI systems continue to improve. The result is valuable precisely because it shows that expertise can make delegation less useful in some settings. The expert may already possess the context the model is struggling to reconstruct.

This is why AI fluency cannot mean maximal AI use. Sometimes the highest-skill move is recognizing that the verification burden will cost more than doing the task directly.

Training Debt Returns

Chapter Six introduced training debt: the organization enjoys the productivity benefit of removing beginner work and later discovers that the beginner work was also training.

The problem becomes sharper as AI improves. If a first-year analyst no longer builds the model, a junior lawyer no longer reads the first set of cases, or a new programmer no longer debugs routine failures, the institution still gets the deliverable. What it may lose is the sequence of mistakes and corrections through which judgment develops.

PwC’s 2026 Global AI Jobs Barometer finds signs that the skill structure is already changing. In U.S. entry-level jobs most exposed to AI, postings were seven times more likely to demand skills traditionally associated with senior workers, such as judgment and leadership. Those more seniorized entry-level roles grew 35% from 2019, while other entry-level roles declined 10%.

PwC’s analysis is observational and spans a period with many labor-market forces, so it cannot attribute every change in those postings to AI. The pattern nevertheless fits the training-debt problem exactly: employers can begin asking juniors to arrive with capabilities that were once developed by doing junior work.

A career ladder fails when the first rung disappears but the second rung still requires experience from the first.

The Outage Test

The customer-support study contains a useful institutional idea hidden inside a technical accident. Researchers could observe workers during periods when the AI system was unavailable. That made it possible to ask a question every AI-using organization should care about: what capability remains when assistance disappears?

Organizations do not need to wait for an outage to find out. They can deliberately test independent performance on selected tasks, rotate people through work that preserves diagnostic skill, require original analysis before showing an AI recommendation, or ask senior reviewers to explain why an answer is right rather than merely approve it. The point is not to stage productivity theater. It is to measure whether competence is being transferred to people or merely rented from the tool.

Aviation turned this logic into operating practice long before generative AI. In Safety Alert for Operators 13002, issued in January 2013, the Federal Aviation Administration urged operators to promote manual-flight opportunities so pilots retain the skills to fly when automation fails or is unavailable. The point was competence maintenance: a mature automated system periodically creates opportunities to practice without the automation because retained human capability is itself part of the safety system.

The appropriate frequency depends on the consequence. A marketing team may tolerate substantial dependence because failure is cheap and reversible. A physician, engineer, lawyer, security analyst, or operator of critical infrastructure may need a much stronger independent baseline. Resilience is part of expertise when the system can fail, become unavailable, or be confidently wrong.

The outage test is therefore a maintenance practice rather than a new Accountability Chain instrument. It asks whether the human side of the system still contains enough knowledge to take over, challenge the machine, and recover when normal delegation stops.

Expertise Is Narrowing and Deepening at the Same Time

The labor market may be producing something stranger than simple deskilling. A 2026 IZA discussion paper used longitudinal O*NET data from 2011 through 2025 to examine how AI exposure relates to the human capabilities jobs require. The authors found that more AI-exposed occupations showed greater proficiency requirements for selected capabilities while also showing a narrower overall breadth of capabilities.

That is consistent with a world in which professionals stop doing some routine components of a job and become more specialized in the parts that remain. The accountant may calculate less and advise more. The developer may type less code and spend more time on architecture. The researcher may run fewer manual experiments and spend more time deciding which experiment deserves scarce compute.

PwC describes a related split between roles that AI professionalizes and roles it democratizes. In the professionalized track, routine work falls away while judgment, creativity, leadership, and domain expertise become more valuable. In the democratized track, tasks become easier for non-experts to perform.

Both can happen at once. AI can lower the expertise needed to produce an acceptable answer while raising the expertise needed to decide which acceptable answer should be trusted.

The Research Intern Is the Expertise Problem in Miniature

The most revealing example may be inside AI research itself. On September 6, 2026, OpenAI said it had reached its self-imposed milestone of an automated research intern. The company defines that narrowly: a system able to carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days. OpenAI also reported that by mid-August its research organization was using 3.1 agent-workdays of coding-agent effort for every human workday.

The numbers come from OpenAI’s own internal measurement. The company defined the milestone, measured the work, and published the result. OpenAI systems are also used in this Living Edition’s production workflow, so the evidence belongs on the page with the conflict visible rather than disguised as an independent benchmark.

The details are more interesting than the headline. OpenAI says people still set research priorities, decide which ideas to pursue, judge results, and decide whether to scale, pause, or deploy. More than half of successful four-to-eight-hour agent tasks in its recent data still required at least one human intervention. High-level planning remained a minimal share of agent output.

This is the expertise problem in miniature. When a machine performs more of the research work, people must still choose worthwhile questions, guide the work, interpret the results, and decide when to stop. Those decisions may produce fewer visible pages or files, but they determine whether the faster work is worth doing.

The faster the machine can do research, the more consequential the human decision about what research should be done.

Current Evidence: Judgment Is Becoming the Scarce Input

Reviewed September 11, 2026. Contracting-workflow evidence added October 7, 2026.

The customer-support evidence shows both sides of the expertise transition. Less experienced workers gained dramatically and learned some behaviors that survived outages. The strongest workers, however, increased adherence even while some quality measures declined, raising the possibility that the people who generate the best training examples may contribute fewer original solutions over time.

The BCG experiment shows why expertise cannot be reduced to prompt technique. AI improved speed and quality across a broad set of consulting tasks, then made consultants substantially less likely to reach the correct conclusion on a task outside the model’s capability frontier. The scarce skill was recognizing the boundary.

Software evidence points in both directions. Three field experiments at Microsoft, Accenture, and a Fortune 100 company covering 4,867 developers found about a 26% increase in completed tasks with an AI coding assistant, with less experienced developers showing greater adoption and gains. METR’s smaller trial of experienced open-source developers working in familiar repositories found a 19% slowdown. The settings differ, and the findings should not be forced into one universal productivity number.

Knowledge workers themselves report a similar shift. The Microsoft and Carnegie Mellon survey found that greater confidence in AI was associated with less critical thinking, while critical thinking increasingly moved toward verification, integration, and task stewardship. The result is self-reported rather than a direct test of long-run skill retention, but it identifies the new location of the cognitive work.

Labor-market evidence is beginning to show the institutional consequence. PwC’s 2026 analysis finds a two-track market in which some AI-exposed jobs become more professionalized and more dependent on judgment, while some are democratized. Entry-level roles most exposed to AI are increasingly asking for traditionally senior skills. The labor market can demand judgment faster than organizations learn how to produce it.

OpenAI’s internal research data pushes the pattern to the frontier. The company says coding agents now contribute multiple agent-workdays for each human research workday and can perform defined multi-day research tasks under supervision. Yet human researchers still set priorities, evaluate results, and intervene frequently on longer tasks. Automation is climbing the ladder while leaving the highest-leverage decisions at the top.

On October 6, 2026, OpenAI reported an evaluation of agents configuring contracting workflows in Ironclad. Across eleven research tasks, GPT-6 Astra averaged 55.0% on the evaluation criteria, compared with 41.6% for GPT-5.6 Sol. These were company-reported rubric scores, not percentages of complete workflows safely finished. The tasks included setting up approval routes and rules for contract terms. Improvement is useful, but an average score cannot establish that every consequential rule works.

Imagine an agent building a purchasing process. A reviewer should test whether requests above the spending limit reach the required approver and whether exceptions follow the intended route. Inspecting the screen is only the beginning. Expertise must check the behavior of the finished process before people rely on it. The authority to configure a rule deserves review because that rule may govern many later decisions.

A September 2026 working paper reports a three-month randomized study involving 133 patent lawyers at eleven U.S. firms. Access to an AI drafting assistant improved assisted work, with larger gains among junior lawyers. On a separate review task participants were instructed to complete without AI, senior lawyers given access outperformed controls, while junior lawyers showed no average advantage. Google funded the experiment’s direct costs. The findings concern one professional setting, one tool and a limited period.

A stronger deliverable therefore does not establish that the person producing it has acquired stronger judgment. Firms should evaluate what junior professionals can explain, detect and correct without assistance, and preserve opportunities to practice those skills.

On November 6, 2026, OpenAI and Ligo introduced BioBasicsBench, a benchmark of 100 tasks using raw laboratory evidence. OpenAI reported that GPT-6 Astra earned an average 66.7 percent of the available grading credit. That is not the proportion of complete experiments performed correctly. The company also described errors involving experimental controls and contradictory observations. These are developer-reported evaluation results, not independent proof of reliable laboratory practice.

The reviewer therefore needs access to the evidence beneath the explanation. Before an AI interpretation determines the next experiment, someone with the relevant expertise should check the raw observations, the controls used to establish whether the experiment worked, and any result that conflicts with the proposed conclusion. When the evidence is insufficient, stopping or repeating the experiment may be the competent decision. A fluent interpretation should not turn an uncertain observation into permission to proceed.

The current evidence therefore does not support a simple story of expertise disappearing. It supports a transfer: routine execution moves toward machines, some novices advance faster, some experts gain extraordinary leverage, and the value of calibration rises wherever the system can be wrong without looking wrong.

Accountability Analysis

Principles 2, 4, 5, and 8 carry most of this chapter. Principle 2 separates capability from permission. A system may be able to perform a professional task without making it wise to delegate every instance of that task, especially when the human is still building the competence needed to check it.

Principle 4 asks whether review is real. Expertise becomes the substance of review. A person who lacks enough domain knowledge to disagree is not a meaningful checker merely because a workflow requires a click.

Principle 5 moves responsibility upstream into work design. Which tasks remain human? Which tasks are used for training? Who gets practice on edge cases? What incentives reward original solutions rather than passive adherence? Those decisions determine whether an organization is consuming expertise or replenishing it.

Principle 8 is the center of the chapter. Accountability depends on competence, and competence decays when the work that maintains it disappears. Institutions therefore have to preserve practice, testing, rotation, mentorship, and occasional tool-free performance where independent capability matters.

The expertise problem is ultimately a renewal problem. A system that uses human judgment faster than it creates new human judgment is spending down an invisible asset.

Practice the outage test

Choose a familiar task that is safe to practice without AI, using a training example rather than a live decision with serious consequences. Work through it independently, then compare your reasoning with a checked example or an experienced colleague’s assessment. Identify one skill to practice. The outage test asks what capability remains when assistance is unavailable.

Accountability Verdict

Organizations should use AI aggressively to accelerate learning, expand capability, and remove work that has little developmental value. Preserving busywork merely because earlier generations suffered through it would be an odd theory of civilization.

They should also identify which tasks create the judgment the organization will still need five years from now. Those tasks may need to remain partly human even when AI could complete them faster, or they may need redesigned training exercises, rotations, simulations, audits, and deliberate periods of independent work.

The practical test is simple: remove the tool. What can the person still do? Can the professional recognize an edge case, reconstruct the reasoning, challenge the output, and produce a competent answer when the system is unavailable or wrong? If not, the organization has productivity without resilience.

AI can make expertise more powerful. It can also make expertise harder to grow.

Chapter Nineteen takes the verification problem outside the workplace. When synthetic text, images, audio, and video become cheap, the scarce skill is no longer only professional judgment. It is knowing what evidence deserves belief at all.

Sources and Notes

1. Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, “Generative AI at Work,” The Quarterly Journal of Economics 140, no. 2 (May 2025): 889-942, published February 4, 2025. https://academic.oup.com/qje/article/140/2/889/7990658. Accessed September 11, 2026. The staggered deployment covered 5,172 customer-support agents. AI assistance increased issues resolved per hour by 15% on average and by about 30% for the least-skilled workers; agents with two months of tenure and AI assistance performed about as well as untreated agents with more than six months. The paper also reports small quality declines for top workers, increasing adherence, evidence of worker learning during outages, and a concern that reduced original contributions from top workers could lower the diversity and quality of future training data. Source record: DA-CH18-S1. Archive reference: pending.

2. Fabrizio Dell’Acqua, Edward McFowland III, Ethan Mollick, Hila Lifshitz, Katherine C. Kellogg, Saran Rajendran, Lisa Krayer, François Candelon, and Karim R. Lakhani, “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality,” Organization Science 37 (2026), published online March 11, 2026. https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838. Accessed September 11, 2026. In a preregistered experiment involving 758 BCG consultants, AI users completed 12.2% more tasks and worked 25.1% faster across 18 tasks inside the AI capability frontier while producing higher-quality work. On a complex task selected to be outside the frontier, AI users were 19 percentage points less likely to produce the correct solution. The study used GPT-4. OpenAI systems are also used in this Living Edition’s production workflow; see the transparency statement. Source record: DA-CH18-S2. Archive reference: pending.

3. Joel Becker, Nate Rush, Beth Barnes, and David Rein, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity,” METR, July 10, 2025. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/. Accessed September 11, 2026. Sixteen experienced open-source developers completed 246 randomized tasks in mature repositories where they averaged about five years of prior experience. Allowing early-2025 AI tools increased measured task-completion time by 19%, even though developers predicted a 24% speedup beforehand and believed afterward that AI had sped them up by about 20%. The authors explicitly caution against generalizing the result to all developers, domains, or future models. Source record: DA-CH18-S3. Archive reference: pending.

4. Hao-Ping Lee, Advait Sarkar, Lev Tankelevitch, Ian Drosos, Sean Rintel, Richard Banks, and Nicholas Wilson, “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers,” Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, Article 1121. https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/. Accessed September 11, 2026. The study surveyed 319 knowledge workers who supplied 936 first-hand examples of AI use. Higher confidence in AI was associated with less reported critical thinking, while higher self-confidence was associated with more. Respondents described critical thinking shifting toward verification, response integration, and task stewardship. The study relies on self-report and does not directly measure long-run skill loss. Source record: DA-CH18-S4. Archive reference: pending.

5. Zheyuan Cui, Mert Demirer, Sonia Jaffe, Leon Musolff, Sida Peng, and Tobias Salz, “The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers,” Microsoft Research, June 2025. https://www.microsoft.com/en-us/research/publication/the-effects-of-generative-ai-on-high-skilled-work-evidence-from-three-field-experiments-with-software-developers/. Accessed September 11, 2026. Randomized field experiments at Microsoft, Accenture, and an anonymous Fortune 100 company covered 4,867 developers. Pooled results show a 26.08% increase in completed tasks for developers with access to an AI coding assistant, with less experienced developers showing higher adoption and larger gains. The paper is a preprint and the result should be read alongside studies from other software settings, including METR’s experienced-developer trial. Source record: DA-CH18-S5. Archive reference: pending.

6. PwC, “2026 Global AI Jobs Barometer: Two futures for jobs in an AI era,” June 15, 2026. https://www.pwc.com/gx/en/1/services/ai/ai-jobs-barometer.rhs.html and https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-jobs-barometer.html. Accessed September 11, 2026. PwC analyzed more than one billion job advertisements across 27 countries and territories and describes a two-track market of professionalized and democratized roles. In 2.4 million U.S. entry-level postings, highly AI-exposed roles were seven times more likely to require traditionally senior human-intensive skills such as judgment and leadership; those roles grew 35% since 2019 while other entry-level roles declined 10%. PwC is a consulting firm with commercial interests in AI adoption; the chapter uses the analysis as labor-market evidence rather than as a causal study. Source record: DA-CH18-S6. Archive reference: pending.

7. Michele Cantarella, Giuseppe Molinari, and Chiara Strozzi, “Becoming The Man Without Qualities? Deskilling in the Age of Artificial Intelligence,” IZA Discussion Paper No. 18751, June 2026. https://www.iza.org/publications/dp/18751/becoming-the-man-without-qualities-deskilling-in-the-age-of-artificial-intelligence. Accessed September 11, 2026. Using longitudinal O*NET data for the U.S. labor market from 2011 through 2025, the authors report a dual pattern: greater AI exposure is associated with higher proficiency requirements for selected capabilities within occupations and with a compression in the overall breadth of capabilities required across more exposed occupations. The paper is a discussion paper and identifies associations rather than a simple causal effect of AI adoption. Source record: DA-CH18-S7. Archive reference: pending.

8. OpenAI, “Research acceleration: The view inside OpenAI,” September 6, 2026. https://openai.com/index/research-acceleration-view-inside-openai/. Accessed September 11, 2026. OpenAI says it has reached its internally defined automated research-intern milestone, meaning a system that can perform well-defined research tasks under human direction that would take a skilled researcher a few days. The company reports 3.1 agent-workdays of coding-agent effort per human research workday by mid-August 2026, while stating that people still set priorities, evaluate results, and make scale, pause, and deployment decisions. OpenAI also reports that more than half of successful four-to-eight-hour agent tasks required at least one human intervention and that high-level planning remained a minimal fraction of agent output. These are OpenAI’s own measurements and definitions, not an independent benchmark. OpenAI systems are used in this Living Edition’s production workflow; see the transparency statement. Source record: DA-CH18-S8. Archive reference: pending.

9. Microsoft Research, “New Future of Work Report 2025,” published 2025 and current online edition. https://www.microsoft.com/en-us/research/wp-content/uploads/2025/12/New-Future-Of-Work-Report-2025.pdf. Accessed September 11, 2026. The report synthesizes research suggesting that generative AI shifts effort from doing toward choosing among outputs and can create deskilling risk when changes in work are not accompanied by upskilling and reskilling. The chapter uses it as synthesis rather than as an independent causal estimate. Microsoft develops and sells AI products and is also an employer represented in several studies cited in the chapter. Source record: DA-CH18-S9. Archive reference: pending.

10. Federal Aviation Administration, Safety Alert for Operators 13002, “Manual Flight Operations,” January 4, 2013. https://www.faa.gov/other_visit/aviation_industry/airline_operators/airline_safety/safo/all_safos. Accessed September 11, 2026. The FAA urged operators to promote manual-flight opportunities in operations and training so pilots maintain proficiency when automation is unavailable or inappropriate. The chapter uses the alert as a precedent for the outage test as a competence-maintenance practice, not as evidence that aviation and generative-AI work are identical risk domains. Source record: DA-CH18-S10. Archive reference: pending.

11. Steven Randazzo, Akshita Joshi, Katherine C. Kellogg, Hila Lifshitz, Fabrizio Dell’Acqua, and Karim R. Lakhani, “GenAI as a Power Persuader: How Professionals Get Persuasion Bombed When They Attempt to Validate LLMs,” Harvard Business School Working Paper 26-021, 2025. https://www.hbs.edu/ris/download.aspx?name=26-021.pdf. Accessed September 11, 2026. In a study involving more than seventy BCG consultants using GPT-4, the authors report that the model increased persuasive tactics as professionals attempted to validate its output through fact-checking, pushback, and exposing inconsistencies. The paper remains a working paper. GPT-4 is an OpenAI system, and OpenAI systems are also used in this Living Edition’s production workflow; see the transparency statement. Source record: DA-CH18-S11. Archive reference: pending.

12. OpenAI, “Advancing computer use with Ironclad,” October 6, 2026. https://openai.com/index/advancing-computer-use-with-ironclad/ Accessed October 7, 2026. Provider research report, not independent validation. Eleven tasks, 8–50 evaluation criteria per task; mean rubric scores 55.0% versus 41.6%, not complete-workflow success rates. The report’s time estimates are simulated, not measured customer savings. Training/evaluation used simulated tasks derived from filtered public SEC contracts, according to OpenAI. The purchasing example and review recommendation are the author’s analysis. OpenAI systems are used in this Living Edition’s production workflow; the standing disclosure remains applicable. Source record: DA-CH18-S12. Update 072. Archive reference: pending.

13. David Autor, Tanya Rodchenko, Josh Martin, Zanna Iscenko, Scott Strand, David Pearl, and Melissa Ferere, “Does AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent Drafting,” NBER Working Paper 35720, September 2026. https://doi.org/10.3386/w35720 and https://www.nber.org/papers/w35720. Authors' explanation: https://research.google/blog/does-better-work-always-mean-better-workers/ October 7, 2026. Accessed October 11, 2026. Registered randomized experiment; working paper, not peer-reviewed publication. Google funded direct experimental costs; limited sample, duration and tool. Primary abstract and authors' explanation checked; full-paper retrieval was unavailable, so no detailed compliance, attrition or robustness claim is made. Task instructions are distinguished from verified absence of AI use. The institutional recommendation is the author's analysis. Source record: DA-CH18-S13. Update 086. Archive reference: pending.

14. OpenAI, “Measuring how AI agents handle everyday wet-lab science,” November 6, 2026. https://openai.com/index/measuring-how-ai-agents-handle-wet-lab-science/ Accessed November 7, 2026. BioBasicsBench was developed with Ligo; developer-reported rubric evaluation, not independent field validation or a measure of complete experimental success. Review recommendations are author analysis. OpenAI systems are used in this Living Edition’s production workflow; see the transparency statement. Source record: DA-CH18-S14. Update 134. Archive reference: pending.