AI is Not Making Expertise Obsolete

From our Political Economy Correspondent: Why the AI Jobs Panic Is Asking the Wrong Question

Guerre and Shalom is so pleased to announce that Oded Faran is our new Political Economy Correspondent, producing exclusive articles on the political-economic questions of our time. Today he joins our symposium on AI, specifically speaking to its effect on the jobs market. This follows in the wake of Democratic Renewal Correspondent Ashok Panikkar’s piece on the same topic, linked below and made reference to by Oded on several occasions in his article.

Oded’s exclusive piece may well be the best written and most influential of the near-800 pieces we have have published on this journal. It’s definitely the best sourced! Accepted that it isn’t an easy read, but it’s all the more rewarding for its dense, comprehensive analysis of the question of the age. It deserves to be read far and wide.

Precisely because of its non-clickbait nature (but its clear importance), I would urge readers to like, restack and share the article even before they’ve read it so as to ensure it goes viral. There is nothing in it one would “disagree” with --it’s not that kind of piece.

I particularly like Oded’s analysis of AI’s effect on the legal profession. With lawyers needing far leaner firms, and initially racking in sky-high profits, competition will naturally ensue. This, in turn, will lead to large cuts in fees and, finally, accessible justice for all. For as long as we can remember, the law has been out of reach, even for Middle Class professionals, making pretensions of the rule of law seem illusory. Yet with AI reducing the price of cognition, Ted the Electrician and Daniel the Writer will be able to take the powerful to court and win. Quite the prospect. - Daniel Clarke-Serret (Editor, Guerre and Shalom)

---

Ashok Panikkar began our AI symposium:

The Most Dangerous Thing about AI

---

AI Is Not Making Expertise Obsolete

Why the AI Jobs Panic Is Asking the Wrong Question

We made production cheap, then transport, then energy, then calculation. Now we are making a good deal of cognition cheap. Which jobs that destroys is a fair question. Who may collect the dividend is probably the better one.

By Oded J.K. Faran (Jacob’s Ledger)

Image retrieved from https://sproutlycareers.com/top-15-jobs-most-threatened-by-ai-emergence/

A familiar question about an unfamiliar input

Guerre and Shalom recently published an essay by Ashok Panikkar that asks what may be the most uncomfortable question about artificial intelligence: what happens to us when we remove the struggle through which people learn to think, judge and become competent?[1] His concern is wider than employment, but one part of it has stayed with me. The boring work that firms are increasingly tempted to hand to AI was often the work through which juniors became seniors.

I think that concern deserves to be taken seriously, and I will come back to it. But I want to approach the same change from another direction, through prices.

Every few generations something that used to be expensive becomes cheap, and the people living through it tend to describe the machine rather than the price change. The power loom, the tractor, the dynamo, the mainframe. The machine is what you can see. The price change is what rearranges the economy. Mechanized production made cloth and iron cheap. Railways and the motor car made movement cheap. Electricity made usable power cheap, and available in small portions wherever a wire could reach. Computers made arithmetic and storage so cheap that we stopped thinking of them as costs at all. (Nobody budgets for multiplication.)

Generative AI belongs in that sequence, with one obvious difference. What it cheapens is part of what we used to call educated thought: drafting, summarizing, translating, retrieving, classifying, writing serviceable code, producing a first-pass analysis of a document nobody wanted to read. For most of modern history you paid a trained person to do these things, by the hour, and the hours were long.

It is a remarkable change, and I understand why people reach for the word unprecedented. But the first question it raises is an old one. What happens when an important input becomes much cheaper? In general form, we know. The price of things made with it falls, and demand for them usually rises. Some of the people who supplied the input lose income, while some of those who combine it with other skills gain. Firms reorganize around the new costs, more slowly than anyone expects. And the gains go to whoever owns what and can bargain, which is the part nobody mentions at the product launch.

Asked that way, the debate over whether AI will “take our jobs” starts to look badly framed. A model passes a bar exam[2] or builds an app from a paragraph of instructions, and the inference follows by reflex: lawyers and programmers are finished. The demonstration is treated as the proof.

My own view is narrower than the optimists would like and less dramatic than the pessimists expect. AI is plainly technologically unfamiliar. Whether its economics are unprecedented is a question of evidence, and permanent mass unemployment is a hypothesis that someone has to demonstrate; it does not follow from watching a chatbot do something clever. I should admit at the outset that this argument did not survive the research intact. It came out narrower than it went in, and it is weakest exactly where Panikkar’s worry points: the bottom rung of professional careers.

A job, after all, is a collection of tasks held together by an employment contract, a desk and a good deal of habit. A junior legal translator renders text, checks terminology against the client’s earlier filings, rebuilds tables that did not survive conversion, flags errors in the source and calls the client when the original says something the client almost certainly did not mean. Technology has no obligation to automate that bundle in the proportions in which it was assembled. Daron Acemoglu and Pascual Restrepo built a whole framework on this: automation pushes workers out of some tasks, new tasks pull them back in elsewhere, and nothing guarantees that the second keeps pace with the first.[3]

Once you think in tasks, the slide from “AI can do part of this job“ to “this job will disappear“ turns out to have a lot of steps in it, most of which are usually skipped. Whether anyone actually uses the tool for those tasks, whether it replaces the worker or makes her faster, whether the firm hires fewer people or just different ones: each is a separate empirical question. Exposure measures get read as forecasts of job loss when they are closer to a map of places where something might happen. The International Labour Organization’s refined exposure index finds that about one worker in four globally is in an occupation with some exposure to generative AI, and only 3.3 percent of employment falls into its highest band.[4] Its own conclusion is that most jobs are more likely to be transformed than eliminated. Exposure is not substitution, substitution is not the death of an occupation, and the death of an occupation is not mass unemployment.

The History

The lazy inductive-historical argument says that people panicked about the loom, the tractor and the computer, the panics were wrong, so this one is wrong too. While there’s some merit in it, on the whole, I do not find that persuasive. The worriers were often right about particular workers, towns and decades. Where they went wrong was in the leap from those losses to a permanent shortage of work for the economy as a whole.

The motor vehicle is the case I find most instructive. The internal combustion engine made motive power cheap. In 1920 American farms held about 25.2 million horses and mules; by 1959, according to the Census of Agriculture, fewer than 3 million remained.[5] The tractor replaced the horse, and also its lunch: the Census Bureau noted that land which had grown feed for work animals was freed for other uses.[6] Teamsters, harness makers and livery stable hands gave way to automobile workers, mechanics and truck drivers, and agriculture, which employed close to two in five American workers in 1900, employed fewer than one in five by 1940 (Figure 1).[7] Many people did not convert, and the movement off the farms was also a story of rural decline.

Figure 1. Agricultural and non-agricultural workers in the United States, 1820–1940. Agricultural employment rose in absolute numbers until 1910, then fell; its share fell throughout. Source: U.S. Census Bureau, Historical Statistics of the United States, 1789–1945, Series D 1–7.[7]

The horse and wagon probably supplies the best pessimistic analogy in the whole debate. Wassily Leontief made it in 1983: horse employment collapsed and stayed collapsed, because the horse had nothing else the new economy wanted.[8] We need to understand that the only serious question about AI is whether human workers are in the position of the teamster or of the horse. The teamster had hands, judgment and a driving license waiting to happen; the horse had only pulling power, and pulling power had become cheap. (A later rejoinder adds two differences that matter: people can own capital, and they can vote.[9]) For the last century, people have been teamsters. The pessimist’s claim is that a general enough cognitive technology might eventually leave them nothing else to offer. That claim is coherent, and it is about the future.

Timing is yet another factor less discussed. Paul David showed in 1990 that electric motors supplied less than 5 percent of the mechanical drive in American factories in 1899, and that it took roughly another two decades to pass half. For much of that time owners swapped the steam engine for a dynamo and changed nothing else, and productivity barely moved; the gains came in the 1920s, some four decades after the first central power stations, when factories were rebuilt around small motors on individual machines.[10] I suspect that today’s measurements of firms bolting a chatbot onto an unchanged workflow understate both the eventual gains and the eventual disruption.

Alas; we cannot have a serious tech-related economic discussion without mentioning the computer. As a result of the computer-adoption revolution, the automated teller machine cut the tellers needed in a typical urban branch from about 20 to 13 between 1988 and 2004, but made branches cheap enough that banks opened many more, and teller jobs did not disappear as ATMs spread.[11] A study of American job titles found that roughly 60 percent of employment in 2018 was in titles that did not exist in 1940, though it also carries a warning optimists tend to skip: new work comes mainly from innovations that complement an occupation and from rising demand for what it produces. Automation alone does not create it.[12]

Across more than a century in which output per worker multiplied many times over, the economy kept finding paid work for most adults who wanted it, while cheaper inputs destroyed particular trades and left many of the people in them worse off for the rest of their lives. Permanent economy-wide technological unemployment has not followed from new technology in the cases economists have studied most closely.

There are real reasons to think AI may behave differently from all of this, and I find only two of them hard to dismiss.

The most obvious is speed. Electrification needed power stations and rebuilt factories, and a tractor had to be built and sold one farm at a time. A model can appear on a million desks almost overnight, at a marginal cost that is small and falling. Stanford’s latest AI Index reports that generative AI reached about 53 percent adoption within three years of its mass release, faster than the personal computer or the internet at the same stage.[13] Capabilities are also improving quickly while retraining takes time, so a skill that complements this year’s models may be substituted by the models of the year the worker finishes acquiring it.

Breadth probably matters too, because the tractor displaced one kind of work, while language models touch programming, translation, clerical work, marketing and customer service at once; the International Monetary Fund has estimated that about 60 percent of jobs in advanced economies are exposed in some form.[14] And the capital flowing into the technology is highly concentrated. The AI Index puts private AI investment in the United States at about 286 billion dollars in 2025, roughly 83 percent of the global total.[13] If very few people own the thing that does the work, the assumption that productivity gains eventually reach wages has less to stand on.

The less serious objection is about “where people go“. We already have ample experience with that. Earlier automation pushed workers out of routine physical work toward language, judgment and social skill. AI has entered that territory too, starting, as it happens, with the tasks we used to give to beginners. One formal model of the limiting case has automation running ahead of capital accumulation, so that wages fall, and machines eventually reaching the upper bound of what humans can do, so that wages collapse.[15] If that happens, the labor market as we know it ends, and the horse becomes the right analogy after all. But it would take a technology that does not yet exist.

What would change my mind about the narrower claim, before any such technology arrives, is a sustained decline in the employment and earnings of experienced workers in exposed occupations, or clear evidence that displaced workers stay out of work instead of moving. So far we have witnessed neither.

The evidence

One Stanford payroll study has become difficult to avoid in this debate, and its August 2026 revision, using payroll data to June, finds no widespread, economy-wide job displacement associated with AI.[16] (Danish evidence points the same way: linking adoption surveys to administrative records, researchers there find precise null effects on earnings and hours in eleven exposed occupations two years after ChatGPT, ruling out effects larger than about 2 percent.[17])

But employment among Americans aged 22 to 25 in the most AI-exposed occupations now stands about 19 percent below where it would be had it kept pace with similar workers in less-exposed occupations, up from 15 percent a year earlier. In levels, employment of 22-to-25-year-olds in the two most exposed fifths of occupations fell about 11 percent between November 2022 and June 2026, while in the three least exposed fifths it grew about 10 percent. Experienced workers show no comparable gap. The adjustment runs through fewer hires rather than more firings, and it is concentrated where observed AI use tends to automate tasks; where AI mostly complements workers, employment is flat or rising. In the revised paper the authors also distinguish occupations built on codified knowledge, the kind written down in textbooks and procedures, where young workers’ employment has fallen, from those built on tacit knowledge acquired through practice and mentorship, where experienced workers’ employment has risen (Figure 2).[16]

Figure 2. Employment change for workers aged 22 to 25, November 2022 to June 2026, by AI exposure of occupation. The authors describe the pattern as descriptive, not causal; older workers show no comparable gap. Source: Brynjolfsson, Chandar and Chen, Stanford Digital Economy Lab, August 2026.[16]

I want to be careful with these numbers, because the authors insist on calling their results descriptive and refrain from assuming causality. The gaps shrink when education is accounted for, some differential trends predate generative AI, and the gaps are larger in the payroll sample than in national surveys.[16] Two economists writing for the Economic Innovation Group make a sharper point: a cohort of 22-to-25-year-olds is replenished almost entirely by hiring, so any fall in graduate hiring shrinks it fast, and graduate hiring had begun falling in 2022, when interest rates rose.[18] The Stanford authors reply that the divergence survives controls for interest-rate exposure, persists outside technology firms, and kept widening long after rates peaked. They still decline to call it proof, and so do I. A separate Harvard study, using résumé data on about 65 million workers at more than 280,000 firms, finds the same pattern inside firms that adopted AI: junior employment falling, senior employment not.[19]

The productivity evidence points the same way from the other side. In a large field study of customer-support agents, an AI assistant raised productivity by 15 percent on average, by about 30 percent for the less experienced and lower-skilled, and barely at all for the most experienced.[20] The tool seems to capture the habits of the best workers and hand them to beginners, which matters for training. Anthropic’s own usage data, which comes from an interested party and classifies conversations rather than jobs, shows business deployments tilting much more heavily toward delegated or automated workflows than individual use does.[21][22]

AI is plausibly part of the cause of the young workers’ shortfall, though nobody yet knows how large a part, or whether the pattern will spread up the age distribution.

The codified and tacit distinction is, I think, the most interesting thing in the data, because it says quite a lot about what expertise is. In an earlier essay on academic attitudes to AI, I argued that the technology is pulling apart things institutions had bundled for so long that nobody had to define them separately.[23] Professional expertise is the same kind of bundle. A senior lawyer, translator or analyst has been paid partly for information, partly for knowing where to look, and partly for turning both into a competent memo, translation or model. And partly for something harder to name: knowing which question the client should have asked, noticing that a figure does not tie out, recognizing that the confident paragraph in front of you is wrong, and signing the document.

Generative AI has driven down the marginal cost of the first three, in some cases spectacularly. Someone whose market value lay in spending six hours finding what can now be found in six seconds has lost scarcity, no doubt.

The tempting word for what remains is judgment, but it is often used to mean whatever the machine cannot do yet, which makes the argument circular. I would rather define it by what it costs. A decision calls for judgment in the economically relevant sense when the objective cannot be fully specified in advance, when relevant information lies outside the record (the client’s circumstances, the counterparty’s habits, the politics of the room), and when checking the result is expensive while errors are costly and fall on someone identifiable. Where a task can be specified in advance and checked cheaply afterwards, I see no reason to think it stays human. Where it cannot, the person who can specify and check complements the machine, and complements of a cheapened input tend to rise in value. Economists studying machine prediction made a version of this argument years ago.[24][25]

The best evidence for it is a little unsettling. In a field experiment with 758 Boston Consulting Group consultants, those using GPT-4 on tasks inside what the researchers called AI’s jagged frontier completed 12.2 percent more tasks, worked 25.1 percent faster and produced significantly higher-quality work. On a task chosen to fall outside the frontier, those using AI were 19 percentage points less likely to reach the right answer.[26] The machine helped, fluently, in the wrong direction.

We probably should not overstate the comfort in this. AI can codify tacit knowledge, which is what the customer-support researchers think happened with their agents, and the line between what is written down and what is learned by doing moves every year. David Autor goes further and argues that AI’s real opportunity is to let many more workers make some of the decisions now reserved for doctors, lawyers, engineers and professors.[27] If he is right, expertise becomes more widely held, which is good for society and less good for the rents of those who hold it now. Some experts will discover that what they took for their judgment was mostly their access. Expertise survives as an economic category, but the market is changing what it means by the word and which parts of it command a premium.

What about the Juniors?

If judgment is the scarce complement to cheap cognition, Panikkar’s question becomes unavoidable: where does tomorrow’s judgment come from? In his words, “the boring grunt work was how juniors became seniors.”[1] It was where they made mistakes, learned what the textbooks left out and slowly acquired judgment. If AI does that work, how does anyone become an expert, and will we still need them?

I think this is the strongest objection to my argument, and the payroll data supports it to an extent. But it bundles two worries: Society may permanently need fewer experts. Or our way of producing experts may have depended on work that is now disappearing. I am fairly confident about the second and unconvinced by the first.

Consider how apprenticeship has been financed. A law firm hires a young associate and bills her document review, due diligence and first drafts to clients, who pay because the work needs doing. She learns from doing it, and from the partner who marks up her drafts. Accounting firms, consultancies, newsrooms and translation agencies run on versions of the same arrangement. The drudgery paid for the apprenticeship. Make it nearly free, and no client will pay for three associate hours on a review a model does in minutes. The training that used to ride along at no visible cost turns up on the firm’s books as a pure expense.

Gary Becker pointed out in the 1960s that, in a competitive labor market, firms have little reason to bear the cost of general training whose benefits a worker can take to another employer.[28] Professional judgment is about as portable as skills get. That gives each firm an incentive to train fewer juniors than the profession as a whole may need and to hope to hire experienced people later from someone else. If every firm reasons that way, the shortage appears a decade later as a scarcity of mid-career practitioners and a rising price for experience, long after the decisions that caused it. The junior-worker problem may be a training crisis before it is an unemployment theory.

Preserving drudgery for its own sake would be plainly silly. Deliberate training is the obvious precedented alternative. Under zero-flight-time rules, a pilot can earn a rating on a new airliner entirely in a full-motion simulator, but only with substantial prior flying experience, and the first real takeoffs and landings are flown under the supervision of a type-rating instructor.[29] Medicine keeps residents supervised for years.[30] In both cases the training is designed rather than incidental, and someone pays for it.

AI itself could help or hurt here. If the customer support assistant helped novices most, it is also a very patient tutor that could make apprenticeship shorter and cheaper. Used as a substitute, it removes the practice altogether. Panikkar’s point about friction applies: a junior who never drafts the first version never learns to see what is wrong with a draft, which is precisely the skill the consulting experiment found to be scarce. Which use wins will depend on what firms are paid for and what they are required to do.

My answer to Panikkar, then, may seem only partial. Economists are usually better at explaining the general mechanisms, the “how”, than at predicting the concrete “what” and “when”. Those emerge through market processes that cannot be known in advance. That is intellectually unsatisfying, certainly, especially when the obvious question is what people will actually do instead. But nevertheless, let me take a shot at satisfying those unhampered intellectual appetites

Let’s agree that the way we produce experts is changing, perhaps quite sharply. However, I am much less convinced that it is breaking. Junior work is unlikely simply to disappear. Its content will change. AI will become another tool in the junior’s toolbox, just as it will in the senior’s, and some of what we now regard as apprenticeship will migrate into tasks, occupations and perhaps entire professions for which we do not yet have names.

That is difficult to picture because we are much better at seeing the work a new technology makes unnecessary than the work it makes possible. A typist watching the first office computers arrive could see perfectly well which part of her job was under threat. She could hardly have been expected to describe the millions of jobs that would later depend on spreadsheets, databases, software, digital design, online commerce or network administration. Those jobs were not merely difficult to predict. Many of them were difficult even to describe before the technologies, firms and markets around them existed.

We face the same problem now. We can already see the first drafts, document reviews, translations and routine analyses that AI can perform. What we cannot see with anything like the same clarity is the work that will become worthwhile once those things are cheap. That ignorance is not evidence that the work will fail to appear. It is a limitation on our ability to specify in advance what a process of economic discovery will produce.

If technological history is any guide, juniors will still have to learn, make mistakes and acquire judgment. That is different from the lazy claim that past fears were wrong and therefore present fears must be wrong too. History cannot tell us what the new work will be, or guarantee that enough of it will appear. It tells us only that new forms of work have often been hardest to see before the technologies and markets that created them existed. Juniors may therefore learn different things, with different tools, while doing work we do not yet recognize as a profession. The career ladder may be changing. Some of its present rungs may disappear. Others have probably not been built yet.

Who gets the AI dividend?

Suppose a lawyer with AI can now do work that used to take three people. The immediate temptation is to count the two missing lawyers, which tells us almost nothing about where the gain ends up.

At first the partner may simply keep it. Fees do not change, two associates are not hired, and profit rises. That gets harder once clients understand what has happened, and while other firms try to constantly compete with one another. Some will demand lower fees, and cheaper legal work may bring in people who could not previously afford it at all; the Legal Services Corporation found that low-income Americans got no legal help, or not enough, for 92 percent of the civil legal problems that substantially affected them, and about half of those who did not seek help cited cost.[31] A smaller practice may also suddenly be able to compete with a firm whose advantage used to depend on armies of juniors, to the extent that the rules on who may provide legal services let it. Where licensing protects incumbents, more of the dividend stays with them, so the distribution of AI’s gains in law will be decided at least as much by regulators and bar associations as by engineers. It is entirely possible that cheaper legal work ends up meaning more legal work, and more lawyers of a somewhat different kind, the way cheaper branches once meant more bank tellers.

The model providers look like the obvious winners, but so far the evidence is more reassuring than that. An earlier edition of Stanford’s AI Index estimated that the cost of querying a model performing at roughly the level of GPT-3.5 fell from about 20 dollars per million tokens in November 2022 to about 7 cents by October 2024 (Figure 3).[32] Prices falling that fast for a given capability look like competition, and the latest edition puts American consumer surplus from generative AI at roughly 172 billion dollars in 2026, a survey-based figure best read as an order of magnitude.[13] Frontier models are another matter, and nothing guarantees that today’s competition lasts.

Figure 3. Price of querying a model at GPT-3.5’s level (64.8) on the MMLU benchmark. A price for one capability level, not a general price of intelligence. Source: Stanford HAI, AI Index Report 2025, ch. 1 (Epoch AI and Artificial Analysis data).[32]

The part of the dividend that may be worrying is the one paid by people not yet in the room. Software shows it as clearly as law. Cheaper software can expand demand for software, so lower production costs do not by themselves imply fewer programmers in total. Yet the Stanford data shows employment of software developers aged 22 to 25 falling sharply.[13] Demand for what a profession produces can rise while the route into it narrows.

The gains from cheap cognition are likely to be widely spread and individually small: a cheaper contract, a free translation, faster software. The losses, in this first phase, look concentrated and visible: a particular cohort of graduates in particular occupations, who did what they were told, acquired the credentials and found the entrance narrowed. That combination has a poor political record. Research on the China trade shock found strong, though not definitive, evidence that American districts more exposed to Chinese import competition underwent an ideological realignment, growing more polarized in some respects and shifting rightward in others.[33] The AI version may be generational rather than regional, and a cohort of educated, articulate, online young people with worse prospects than their older siblings is not politically inert. That seems to me a more concrete threat to democratic stability than deepfakes, and whether it materializes depends on whether prices fall, whether entry opens, and whether anyone takes responsibility for training the next cohort of experts.

I set out to argue that the AI jobs panic was mistaken. I now am convinced it is mostly asking the wrong question. The evidence does not show AI producing mass unemployment. But something is plainly happening to work, and the first place it shows is, surprisingly, not the unemployment rate.

It shows first in the changing content of work, in who gets hired, in which tasks retain value, and in which forms of expertise become easier to acquire or less scarce. That is a material adjustment that cannot be dismissed, and for some workers it will be painful. But adjustment per se is not equal to obsolescence.

I believe the error is in looking at the work AI can already do and treating that visible subtraction as a forecast of the labor market that follows. We can name the drafts, reviews, translations and analyses becoming cheaper because those jobs already exist. We cannot name with equal confidence the work that becomes possible once cognition itself is cheaper. That asymmetry should make us cautious about predictions of permanent displacement. The typist could see the computer taking over typing long before anyone could describe the occupations that spreadsheets, databases, software and networks would eventually support.

I expect AI to follow the same broad logic, though not necessarily the same path or timetable. Juniors and seniors alike will use it as part of their work. Some old tasks will disappear. Others will become more valuable. Entire occupations may emerge around capabilities we currently regard as mere tools. None of this guarantees a painless transition. History offers no guarantee that every displaced worker will be made whole. It does give us good reason to distinguish between work we can already see disappearing and work whose economic value has not yet been discovered.

That is why we should not focus on whether AI will leave us with fewer jobs. The better question would be: what becomes scarce when cognition becomes cheap, who learns to supply it, and who ultimately receives the dividend?

Sources

[1] Panikkar, Ashok. 2026. “The Most Dangerous Thing about AI.” Guerre and Shalom, 16 September 2026. https://danielclarkeserret.substack.com/p/the-most-dangerous-thing-about-ai

[2] OpenAI. 2023. “GPT-4 Technical Report.” https://openai.com/index/gpt-4-research/

[3] Acemoglu, Daron, and Pascual Restrepo. 2019. “Automation and New Tasks: How Technology Displaces and Reinstates Labor.” Journal of Economic Perspectives 33 (2): 3–30. https://doi.org/10.1257/jep.33.2.3

[4] International Labour Organization. 2025. “Generative AI and Jobs: A Refined Global Index of Occupational Exposure.” ILO Working Paper 140. https://www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-occupational-exposure

[5] U.S. Bureau of the Census. 1962. U.S. Census of Agriculture: 1959, Vol. II, General Report, ch. 7, “Livestock and Livestock Products,” Tables 2, 3 and 18. https://www.nass.usda.gov/AgCensus/archive/files/1959-Livestock_and_Livestock_Products-TABLES-913-Table-02.pdf

[6] U.S. Bureau of the Census. 1956. U.S. Census of Agriculture: 1954, Vol. III, Agriculture 1954: A Graphic Summary, “Horses and Mules.” https://www.nass.usda.gov/AgCensus/archive/files/1954-Agriculture_1954_A_Graphic_Summary-MAPS_AND_CHARTS-1089-Table-32.pdf

[7] U.S. Bureau of the Census. 1949. Historical Statistics of the United States, 1789–1945, ch. D, Series D 1–7. https://www2.census.gov/library/publications/1949/compendia/hist_stats_1789-1945/hist_stats_1789-1945-chD.pdf

[8] Leontief, Wassily. 1983. “Technological Advance, Economic Growth, and the Distribution of Income.” Population and Development Review 9 (3): 403–10. https://doi.org/10.2307/1973315

[9] Brynjolfsson, Erik, and Andrew McAfee. 2015. “Will Humans Go the Way of Horses?” Foreign Affairs 94 (4).

[10] David, Paul A. 1990. “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox.” American Economic Review 80 (2): 355–61.

[11] Bessen, James. 2015. “Toil and Technology.” Finance & Development 52 (1). International Monetary Fund. https://www.imf.org/external/pubs/ft/fandd/2015/03/bessen.htm

[12] Autor, David, Caroline Chin, Anna Salomons, and Bryan Seegmiller. 2024. “New Frontiers: The Origins and Content of New Work, 1940–2018.” Quarterly Journal of Economics 139 (3): 1399–1465. https://doi.org/10.1093/qje/qjae008

[13] Stanford Institute for Human-Centered AI. 2026. AI Index Report 2026, “Economy.” https://hai.stanford.edu/ai-index/2026-ai-index-report/economy

[14] International Monetary Fund. 2024. “Gen-AI: Artificial Intelligence and the Future of Work.” Staff Discussion Note SDN/2024/001.

[15] Korinek, Anton, and Donghyun Suh. 2024. “Scenarios for the Transition to AGI.” NBER Working Paper 32255. https://www.nber.org/papers/w32255

[16] Brynjolfsson, Erik, Bharat Chandar, and Ruyu Chen. 2026. “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence.” Stanford Digital Economy Lab, revised August 2026. https://digitaleconomy.stanford.edu/news/canariesaug26/

[17] Humlum, Anders, and Emilie Vestergaard. 2025, revised March 2026. “Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI.” NBER Working Paper 33777. https://www.nber.org/papers/w33777

[18] Iscenko, Zanna, and Fabien Curto Millet. 2026. “Looking for the Ladder: Is AI Impacting Entry-Level Jobs?” Economic Innovation Group, American Worker Project, January 14, 2026. https://eig.org/wp-content/uploads/2026/01/TAWP-Iscenko-Millet.pdf

[19] Hosseini Maasoum, Seyed Mahdi, and Guy Lichtinger. 2025. “Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data.” SSRN Working Paper 5425555, last revised 6 June 2026. https://doi.org/10.2139/ssrn.5425555

[20] Brynjolfsson, Erik, Danielle Li, and Lindsey Raymond. 2025. “Generative AI at Work.” Quarterly Journal of Economics 140 (2): 889–942. https://doi.org/10.1093/qje/qjae044

[21] Anthropic. 2025. “Anthropic Economic Index Report: Uneven Geographic and Enterprise AI Adoption.” September 2025. https://www.anthropic.com/research/economic-index-geography

[22] Anthropic. 2026. “Anthropic Economic Index Report: Economic Primitives.” January 2026. https://www.anthropic.com/research/anthropic-economic-index-january-2026-report

[23] Faran, Oded J.K. 2026. “What Is Really Behind Academic AI Puritanism? On AI-Assisted Writing, Academic Credentials, and the Nature of Expertise.” Jacob’s Ledger, 9 August 2026.

What Is Really Behind Academic AI Puritanism?

[24] Agrawal, Ajay, Joshua Gans, and Avi Goldfarb. 2018. Prediction Machines: The Simple Economics of Artificial Intelligence. Boston: Harvard Business Review Press.

[25] Agrawal, Ajay, Joshua Gans, and Avi Goldfarb. 2019. “Artificial Intelligence: The Ambiguous Labor Market Impact of Automating Prediction.” Journal of Economic Perspectives 33 (2): 31–50. https://doi.org/10.1257/jep.33.2.31

[26] Dell’Acqua, Fabrizio, Edward McFowland III, Ethan Mollick, Hila Lifshitz, Katherine C. Kellogg, Saran Rajendran, Lisa Krayer, François Candelon, and Karim R. Lakhani. 2026. “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality.” Organization Science 37 (2): 403–423. https://doi.org/10.1287/orsc.2025.21838

[27] Autor, David. 2024. “Applying AI to Rebuild Middle Class Jobs.” NBER Working Paper 32140. https://www.nber.org/papers/w32140

[28] Becker, Gary S. 1964. Human Capital: A Theoretical and Empirical Analysis, with Special Reference to Education. New York: National Bureau of Economic Research; distributed by Columbia University Press.

[29] European Union Aviation Safety Agency. Easy Access Rules for Aircrew (Regulation (EU) No 1178/2011), provisions on zero-flight-time training. https://www.easa.europa.eu/en/document-library/easy-access-rules/online-publications/easy-access-rules-aircrew-regulation-eu-no

[30] Accreditation Council for Graduate Medical Education. Common Program Requirements (Residency), section on supervision. https://www.acgme.org/programs-and-institutions/programs/common-program-requirements/

[31] Legal Services Corporation. 2022. The Justice Gap: The Unmet Civil Legal Needs of Low-income Americans. https://justicegap.lsc.gov/resource/executive-summary

[32] Stanford Institute for Human-Centered AI. 2025. AI Index Report 2025, ch. 1, “Research and Development.” https://hai.stanford.edu/assets/files/hai_ai-index-report-2025_chapter1_final.pdf

[33] Autor, David, David Dorn, Gordon Hanson, and Kaveh Majlesi. 2020. “Importing Political Polarization? The Electoral Consequences of Rising Trade Exposure.” American Economic Review 110 (10): 3139–83. https://doi.org/10.1257/aer.20170011