AI and the Graduate Premium: Knowledge Gets Cheaper, Judgement Gets Dearer

Half a percentage point. That is how much a year of schooling has lost in wage value, in the American occupations most exposed to generative AI, since ChatGPT arrived in November 2022. It sounds like a rounding error. It is not. It is one of the first direct estimates of something economists have suspected for three years: the machine is coming for the part of your education that can be written down.

The number comes from a working paper posted to arXiv on 26 October 2026 by Eduardo Levy Yeyati, Darío Judzik and Juan Ignacio Barbiero of Universidad Torcuato Di Tella in Buenos Aires. It is titled “Practice Makes Perfect: Tacit Knowledge, Generative AI, and Returns to Schooling and Experience”, and it does something deceptively simple. It takes the workhorse equation of labour economics, the one that tells you how much an extra year in a classroom or an extra year on the job adds to your pay, and asks whether its coefficients moved when large language models went mainstream.

They did. In highly exposed occupations, the return to a year of schooling fell by 0.55 log points, roughly 7 per cent of what it had been. The return to experience did something stranger. It did not rise or fall uniformly. It rotated, like a seesaw. Early-career experience became more valuable. Late-career experience became less so. The pivot sits at about 18 years.

Put that next to two other pieces of 2026 research and a picture forms. PwC's Global AI Jobs Barometer, built from more than a billion job advertisements, finds the labour market splitting into two tracks, with the jobs where AI raises the bar for expertise growing twice as fast as the jobs where AI lowers it. And a field experiment led by the MIT economist David Autor finds that when patent lawyers trained alongside an AI assistant were later asked to work without it, the senior lawyers kept their gains while the juniors, on average, did not.

Together, the three studies describe a repricing. Codified knowledge, the stuff a degree certifies, is getting cheaper because a machine can supply it on demand. Tacit judgement, the stuff you only get by doing the work, is getting dearer, but it has a shelf life. For anyone who has just borrowed tens of thousands to buy a credential, that is uncomfortable news. Read carefully, it is also not the catastrophe the headline numbers suggest.

Seven and a Half Points a Year

To understand what moved, you need to know what was being measured. In 1974 the Columbia economist Jacob Mincer published “Schooling, Experience, and Earnings”, a slim book from the National Bureau of Economic Research that gave the discipline its workhorse. Mincer's insight was that if you take the logarithm of a person's earnings, you can model it as a straight-line function of their years of schooling plus a curved function of their years of work experience. The curve is concave: experience pays a lot early, then the gains flatten as people stop investing in new skills.

The schooling coefficient has an intuitive reading. Because the left-hand side is in logs, a coefficient of 0.075 means each additional year of school is associated with roughly 7.5 per cent higher hourly pay, holding experience constant. Economists call that 7.5 log points. Half a century of refinement later, it remains the default lens for asking what education is worth.

This matters for reading the new paper correctly, because the units are easy to garble. Levy Yeyati and his colleagues report that before ChatGPT, the return to a year of schooling in the most AI-exposed occupations was 7.5 log points. In the least exposed occupations it was just 1.7. After the release, the exposed-group figure fell by 0.55 log points relative to the control group. That does not mean wages fell by 0.55 per cent, nor that the degree premium collapsed by 7 per cent of your salary. It means the slope of pay against schooling flattened, from roughly 7.5 points per year to roughly 7 points per year. The 7 per cent figure is the decline in that slope relative to where it started: 0.55 divided by 7.5 is about 7.3 per cent.

A rough illustration helps. A four-year degree in a highly exposed occupation was previously associated with about 30 log points more pay than a high school diploma, all else equal (four years at 7.5 points each). At the post-shock rate, that becomes about 28 points. That arithmetic is mine, not the authors', and it ignores everything else that changed, but it conveys the order of magnitude. The degree still pays. It pays a little less, and in exactly the places where the technology bites hardest.

How a Degree Lost Half a Point

The study draws on the US Current Population Survey's outgoing rotation groups, the monthly household survey that asks a subset of respondents about their earnings. The authors use data from 2018 through June 2026, covering 655,919 observations of people aged 16 to 64 with a valid hourly wage. They date the shock to November 2022 and organise their years from November to October so that the ChatGPT release falls cleanly on a boundary.

To decide which occupations count as exposed, they use an index built by Edward Felten, Manav Raj and Robert Seamans and published in the Strategic Management Journal in 2021. Crucially, that measure predates ChatGPT, so it cannot have been contaminated by what happened afterwards. The paper compares occupations in the top quartile of exposure, which include software developers, financial managers and lawyers, with those in the bottom quartile, which include truck drivers, construction labourers and waiters. The middle two quartiles are set aside.

The core specification is a Mincer equation with the schooling and experience terms interacted with two switches: whether the occupation is exposed, and whether the observation comes after the release. The authors control for a long list of demographic and job characteristics, with fixed effects for month, state and occupation.

One technical choice deserves attention because it affects the result. High earners in the survey are top-coded: above a certain threshold, the Census Bureau records a cap rather than the true figure. That cap changed in April 2023, from a fixed value to a dynamic one, which happens to fall in the post-shock window. Top-coding is concentrated in the treated occupations: 11.5 per cent of that group is censored, against 0.7 per cent of the controls. The authors handle it with interval-censored maximum likelihood rather than imputation, and they test alternatives. Applying the pre-2023 cap throughout gives a schooling effect of minus 0.53 log points, almost identical to the headline. But a deliberately harsh check, which treats dynamically top-coded observations as exact values, shrinks the schooling estimate to minus 0.09. The authors flag this openly. It is the most important caveat in the paper, and anyone quoting the 0.55 figure should know it exists.

Other checks are reassuring. The pre-release schooling coefficients show no trend, and the break appears in the first year after release. A placebo that re-dates the shock to January 2020, to test whether the pandemic is doing the work, finds nothing. Dropping the technology sector “does little”. Swapping the Felten exposure index for the “GPTs are GPTs” measure developed by Tyna Eloundou and colleagues at OpenAI and the University of Pennsylvania gives a schooling decline of 0.42 log points, close to the baseline.

Experience Turns on a Hinge at Eighteen Years

The schooling result is the headline, but the experience result is stranger. In a standard Mincer setup, the value of an extra year of experience depends on how much you already have. A third year adds more than a twenty-third. What Levy Yeyati, Judzik and Barbiero find is that the shape of that curve changed after November 2022 in exposed occupations. At three years of potential experience, the return to an additional year rose by 0.26 log points. At five years, it rose by 0.23. At fifteen, the change is statistically indistinguishable from zero. At twenty-five years, it fell by 0.12. Under the baseline quadratic specification, the effect crosses zero at 18.1 years, with a standard error of 2.1.

The authors are careful about what this means. It is a change in the slope of the experience profile, not a statement that workers with 18 years of experience are the break-even point for winners and losers in pay. “Potential experience” is also a blunt instrument: it is simply age minus years of schooling minus six. It counts time since leaving education, not time spent doing the tasks AI now touches.

The heterogeneity analysis sharpens the picture. Among workers with an associate degree or higher, the early-career return rises by 0.43 log points at three years, and the late-career return falls by 0.12 at twenty-five. Among workers without a credential, the early gain is smaller and the late decline is not statistically significant. In other words, the rotation is strongest precisely among people who hold formal qualifications. The authors write that “wage differences by early-career experience become steeper after the shock” within that credentialed group.

Finally, the level effect. For a worker with 16 years of schooling and 15 years of experience in a treated occupation, the estimated wage change is minus 3.2 log points. That is a real cut, and it falls on exactly the profile of a mid-career graduate professional. But the authors stress that the level effect “is not itself a theoretical prediction”, and they say plainly that “the findings are not a forecast of long-run returns to education or experience.”

Codified Knowledge Was Always the Cheap Part

The theory behind the paper rests on an old distinction. In 1966 the chemist and philosopher Michael Polanyi published “The Tacit Dimension”, which contains the line “we can know more than we can tell.” Some knowledge can be written down, taught in lectures and tested in exams. Some can only be acquired by doing: the feel for when a contract clause will cause trouble, the instinct for which bug report is really an architecture problem, the sense of when a patient's symptoms do not add up.

In 2014 David Autor borrowed Polanyi's line for an NBER working paper titled “Polanyi's Paradox and the Shape of Employment Growth”. His argument was that computers had automated tasks that could be specified as explicit rules, and had struggled with tasks requiring tacit knowledge. That pattern, he argued, explained why employment had grown at the top and bottom of the wage distribution and hollowed out in the middle.

Generative AI scrambles that story. Large language models are not programmed with rules; they learn from enormous quantities of text, absorbing much of what humanity has managed to write down. They are, in effect, machines for supplying codified knowledge at near-zero marginal cost. Ask one to explain the doctrine of equivalents in patent law or the assumptions behind a discounted cash flow model, and it will usually produce a competent answer in seconds.

The Levy Yeyati model formalises three consequences. First, substitution: machine and worker codified knowledge are treated as perfect substitutes, so when the machine version becomes abundant, the worker version, proxied by schooling, loses scarcity value. Second, complementarity: practice-based judgement becomes more valuable when the codified knowledge it works on gets cheaper, steepening the early-career slope. Third, obsolescence: experience tied to the old way of doing the work loses value when the technology changes, flattening the late-career slope.

The paper concedes its estimates do not separately identify these mechanisms. The pattern is consistent with all three acting together, but does not prove it. Nor does the paper distinguish fields of study or institutional quality. A year of organic chemistry and a year of media studies count the same.

A Labour Market Splitting in Two

If the Levy Yeyati paper measures what workers are paid, PwC's 2026 Global AI Jobs Barometer measures what employers are asking for. Released on 15 June 2026, it analyses more than one billion job advertisements across 27 countries and territories on six continents.

Its central finding is a fork. In some occupations, which PwC calls professionalised, AI automates the routine parts of the job and pushes the remaining human work towards judgement and expertise. PwC's examples include radiologists and recruiters. In others, which it calls democratised, AI makes the work easier for non-experts to perform, lowering skill barriers and shifting roles towards less specialised activity. Its examples include IT service managers and medical secretaries.

According to PwC Belgium's release of the global results, “the 22% of jobs that are being professionalised are growing twice as fast as democratised jobs (52% of jobs)”, while “seeing 42% higher wage growth since 2021.”

Read alongside the earnings data, this is striking. The democratised track looks like the substitution story: AI takes over the codified core of a job, and the premium for having learned it shrinks. The professionalised track looks like complementarity: AI clears away routine work and leaves a residue of judgement that commands a premium. That the professionalised track is smaller but faster growing and better paid fits the idea that value is migrating towards tacit expertise.

PwC also reports that the average wage premium for AI skills in job postings rose to 62 per cent, from 57 per cent the previous year. Pete Brown, PwC's Global Workforce Leader, put the implication directly: “Organisations need to rethink how they develop talent if they want people to thrive in this new environment.”

Job advertisements capture what employers say they want, not what they pay individual workers, and PwC sells AI advisory services. Its findings are a large signal about demand, not a peer-reviewed causal estimate.

Junior Jobs That Ask for Senior Skills

The most pointed PwC finding concerns the bottom rung. Analysing 2.4 million entry-level job postings in the United States, the barometer finds that junior roles highly exposed to AI are seven times more likely to require “traditionally senior skills”, such as leadership and strategic thinking, than entry-level roles that are not exposed. Some accounts express this as 52 per cent of new skills against 7 per cent. That ratio fits the sevenfold headline, but the breakdown does not appear in PwC's public summaries, so the sevenfold figure is the safer one to quote.

Openings for these “seniorised” roles have grown 35 per cent since 2019, while other entry-level openings have fallen 10 per cent. The entry-level job is being redefined: employers want new hires who can do on day one what used to take years to learn.

Other data point the same way. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen of Stanford's Digital Economy Lab have been tracking employment using payroll data from ADP, the largest payroll processor in the United States. Their paper “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence”, first released in August 2025 and revised on 12 August 2026 with data through June 2026, finds no evidence of widespread, economy-wide job displacement. But employment of workers aged 22 to 25 in AI-exposed occupations is now 19 per cent below where it would be had it kept pace with less exposed peers. Experienced workers show no comparable gap. The adjustment works mainly through reduced hiring of young people rather than layoffs, and it is concentrated where AI substitutes for human tasks rather than complementing them.

The Federal Reserve Bank of New York's tracker of the labour market for recent college graduates tells a similar story. In the third quarter of 2026, the unemployment rate for recent graduates eased to 5.2 per cent, but the underemployment rate, the share working in jobs that do not typically require a degree, rose to 42.5 per cent.

There is an apparent paradox here. If early-career experience is becoming more valuable, why are young people finding it harder to get hired? The Levy Yeyati paper measures returns among people who are employed. Once you are in the door, each year of practice is worth more than before. But the door is narrower, because the entry-level job, where codified knowledge was traditionally converted into judgement, is exactly what AI is eroding. The authors note that changes in who gets employed could partly explain their early-career estimates. The premium on early experience and the shortage of early jobs may be two faces of the same scarcity.

The Patent Attorneys Who Kept Their Gains

Which brings us to the most direct test of whether AI helps or hinders the acquisition of judgement. In September 2026 the NBER released working paper 35720, “Does AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent Drafting”, by David Autor, Tanya Rodchenko, Josh Martin, Zanna Iscenko, Scott Strand, David Pearl and Melissa Ferere. A revision followed in October.

The design is unusually clean for a workplace study. In a pre-registered randomised controlled trial, 133 practising patent lawyers at eleven US intellectual property firms were split into groups, with roughly two-thirds given access to a custom AI patent drafting assistant. Blinded expert patent attorneys scored all the work. The researchers measured performance with the tool and then, after three months, tested everyone on a core professional task without it: redlining an existing patent application.

With the tool, everyone got better. AI access raised the quality of work on benchmark drafting tasks by 0.34 standard deviations at 10 days and 0.38 at 90 days, with larger gains among junior lawyers. Dennis Crouch, the University of Missouri law professor who writes the Patently-O blog, noted that much of the improvement came from “avoiding awful work” rather than producing the very best applications.

Without the tool, the picture split. Treated lawyers outperformed controls by 0.32 standard deviations on the unaided redlining task. But that advantage came entirely from senior lawyers, whose gain was 0.45 standard deviations. Junior lawyers showed no average improvement. Their scores spread out: mediocre results became much less common, while both poor and good results became more common. The authors describe AI as acting as a “springboard” for some juniors and a “cushion” for others. Their conclusion is a sentence worth pinning above every graduate trainee's desk: “Foundational expertise may be a prerequisite for extracting durable skill from AI-assisted practice.”

Autor summarised the result to Fortune in a phrase: AI is “a performance equalizer, but a skill-disequalizer.” He added a warning aimed squarely at people starting out: “Young professionals need to beware of the illusion of competence.” Junior lawyers given the tool reported a surge in task satisfaction, the largest effect in the study. They enjoyed skipping the blank page and moving straight to a reviewer's role. The danger is that the satisfaction of reviewing a competent draft is not the same as acquiring the ability to produce one.

The caveats are substantial, and Fortune's reporting set them out. The paper is not peer-reviewed. Only 91 of the 133 lawyers completed the AI-free test. The firms declined a baseline skills test, so skill gains are inferred from random assignment rather than measured before and after. Fifteen of the 91 markups were flagged as possibly AI-assisted despite the ban; excluding them shrinks the senior advantage to 0.39 standard deviations at a looser statistical threshold, and the overall effect no longer holds. Five of Autor's co-authors are full-time Google employees and the sixth is a paid Google contractor, Google conducted and paid for the study, and the tool was Google's own. Autor is head of MIT's economics department and a Google Technology and Society Visiting Fellow. Three months is short compared with the years it takes to make a patent lawyer.

Still, the result rhymes with others. Erik Brynjolfsson, Danielle Li and Lindsey Raymond's study of more than 5,000 customer-support agents, published in the Quarterly Journal of Economics in 2025, found that AI assistance raised productivity by 15 per cent on average, with the biggest gains for less experienced and lower-skilled workers, while the most experienced saw small gains in speed and small declines in quality. And a field experiment by Hamsa Bastani and colleagues, published in the Proceedings of the National Academy of Sciences in 2025, found that Turkish high school students given unrestricted access to GPT-4 during maths practice scored 17 per cent lower once the tool was taken away, compared with students who never had it. A version with learning-protective guardrails largely mitigated the harm.

The pattern across all three is consistent. AI lifts novice performance in the moment. Whether it builds novice capability over time depends on what the novice already knows and how the tool is used.

Why Experience Expires

The other half of the rotation, the falling return at 25 years, connects to a literature on skill obsolescence that predates generative AI. In 2020 David Deming of Harvard and Kadeem Noray published “Earnings Dynamics, Changing Job Skills, and STEM Careers” in the Quarterly Journal of Economics. Using job vacancy data from Burning Glass Technologies covering 2007 to 2019, they showed that the earnings premium for graduates in technology-intensive subjects such as computer science, engineering and business declines rapidly as they gain experience.

Their model treats a career as a race between on-the-job learning and skill obsolescence. In fast-changing fields, the specific skills a graduate brings are replaced within a few years, which lowers the return to experience and flattens the age-earnings profile. Higher-ability STEM graduates also tend to leave STEM careers more often and at younger ages.

Generative AI may be turning a phenomenon that was confined to technology-intensive careers into something more general. If a large share of codified knowledge in law, finance, marketing and software is now supplied by a machine, then the experience of having done those tasks the old way may depreciate across the board. The Levy Yeyati paper's late-career finding is consistent with that reading, though the authors place less weight on individual late-career coefficients because the pre-trend test for them is weaker.

Here is the double bind for someone starting out. Early experience matters more than ever, because judgement is what AI cannot yet supply. But that experience may itself be technology-specific, and if the technology keeps shifting, its value will erode too. The patent lawyer study suggests one way out: seniors who already had deep foundations were able to absorb the new tool and come out stronger. Experience did not simply expire. It compounded, for those who had enough of it to build on.

The Debt Was Priced for a Different Market

None of this would matter so much if degrees were cheap. They are not. In the United States, Federal Student Aid reported in September 2026 that 42.3 million recipients hold federal student loans totalling more than $1.7 trillion. More than 9.3 million are in default, owing $234 billion. In England, the Student Loans Company's 2025-26 statistics put the higher education loan balance at £294.6 billion, and the provisional average balance for borrowers entering repayment in 2026 is £47,730. The House of Commons Library projects the outstanding balance could reach around £500 billion, in 2025-26 prices, by the late 2040s.

Those debts were taken on at prices set for a labour market in which the return to schooling was stable. If the premium is now shrinking in the fields most exposed to AI, the implied return on that investment shrinks with it. A loan agreement does not reprice when the asset it financed does.

It is worth being precise about the scale. A decline of roughly 0.55 log points per year of schooling, or 3.2 log points at the paper's illustrative mid-career profile, is a meaningful cut but not an annihilation. Even after the shock, the return to schooling in exposed occupations, roughly 7 points per year, remains far higher than the 1.7 points in the least exposed occupations. A degree is still associated with substantially higher pay. What has changed is the direction of travel and the reason for the premium. The share of the premium attributable to knowing things is falling. The share attributable to being able to do things is rising.

England's income-contingent repayment partly cushions graduates, because repayments fall when earnings do; many American borrowers have less protection. Either way, the risk sits with the person least able to diversify it: a young worker with a single human-capital asset and a large liability attached.

What the Numbers Do Not Say

First, the Levy Yeyati paper is a working paper, not yet peer-reviewed, and its headline schooling estimate is sensitive to how top-coded wages are treated. Its central identifying assumption, that greater exposure means a larger expansion of machine-supplied codified knowledge, is assumed rather than tested. The authors list possible confounds candidly: changes in who is employed, signalling and employer learning, and post-2022 shifts in worker composition.

Second, years of schooling are a crude proxy for codified knowledge. A degree also signals persistence, conscientiousness and the ability to learn. It provides networks, and it often functions as a legal or professional gatekeeper. None of those functions is obviously diminished by a chatbot, and the paper cannot separate them.

Third, each of the other studies has its own blind spots, from the Autor experiment's single tool to the descriptive nature of the payroll data. Their convergence is more persuasive than any one result.

Fourth, all of this describes the first three and a half years of a technology that is still changing quickly. The Levy Yeyati authors call their results a short-run repricing of existing stocks of human capital. Over the longer run, schooling choices, curricula, mentoring, occupational sorting and the organisation of entry-level work will all adjust. They close by asking “whether workers can continue to acquire the practice-based knowledge” that AI increasingly substitutes for in learning itself.

Building Judgement Without the Apprenticeship

So how should someone at the outset of a career think about a credential? The evidence suggests five principles, none of them comfortable.

Treat the degree as a foundation, not a product. Senior patent lawyers extracted durable skill from AI because they had foundations to build on, and the authors' conclusion that foundational expertise “may be a prerequisite” is an argument for deep formal education, not against it. What loses value is the degree as a certificate of facts possessed. What retains value is the degree as scaffolding for judgement. The goal is not to memorise what a model can recite, but to understand a field well enough to know when the model is wrong.

Test yourself without the tool. Autor's advice to Fortune was that the only reliable check on skill development is performing tasks without AI and evaluating the results. The Bastani study points the same way: students who used unrestricted GPT-4 as a crutch did worse once it was gone. If your employer gives you an AI assistant, the most valuable thing you can do for your own career may be to periodically do the work without it and compare.

Hunt for the practice, not just the job. Returns to early-career experience are rising while early-career hiring is falling. The scarce resource is not the qualification but the chance to convert it into judgement through supervised real work. Roles offering genuine apprenticeship, with feedback from experienced people on tasks that matter, are worth more than their starting salaries suggest.

Lean towards the professionalised track. Ask whether AI in a given role raises the bar for human expertise or lowers it. Roles where the machine handles the routine and leaves you the hard calls are growing faster and paying better. The question applies to specific jobs and employers, not just titles.

Plan for your experience to expire. Technology-specific know-how depreciates, so the defence is to keep renewing it, layering new techniques onto durable understanding, as the senior patent lawyers did, rather than swapping one set of specific skills for another.

For employers, the implication is sharper. Cutting entry-level hiring because AI can do the junior work is a decision to stop producing senior people. The juniors whose outcomes spread out in both directions in the Autor experiment are next decade's senior lawyers, or not.

A Credential Worth Holding, Priced Differently

What does it mean when the wage premium for education shrinks in the fields most exposed to the technology? The honest answer is that the premium was always paying for two things at once, and the market has started to separate them.

One part paid for possessing codified knowledge, and that part is falling, because a machine can now produce it in seconds. The other part paid for the capacity to learn, judge and build on that knowledge, and that part is holding up, perhaps rising, for those who get the chance to exercise it. The degree you paid for is not worthless. Its price has been revised, and the reason it is valuable has shifted from what it certifies you know to what it enables you to become.

For the worker carrying the debt, that is cold comfort, because the debt did not reprice. For the person deciding whether to start, it is a clearer guide than the noise suggests. The credential still pays, most of all to those who treat it as the start of an apprenticeship rather than the end of an education. Half a percentage point is a warning, not a verdict. The 18-year hinge is a reminder that judgement, too, has a shelf life, and the only defence is to keep building it.

References and Sources

  1. Levy Yeyati, Eduardo; Judzik, Darío; Barbiero, Juan Ignacio. “Practice Makes Perfect: Tacit Knowledge, Generative AI, and Returns to Schooling and Experience.” arXiv:2610.36008 [econ.GN], submitted 26 October 2026. Also circulated as Universidad Torcuato Di Tella, Escuela de Gobierno, Documento de Trabajo 2026/15. https://arxiv.org/abs/2610.36008 (HTML full text: https://arxiv.org/html/2610.36008)
  2. Autor, David; Rodchenko, Tanya; Martin, Josh; Iscenko, Zanna; Strand, Scott; Pearl, David; Ferere, Melissa. “Does AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent Drafting.” NBER Working Paper No. 35720, September 2026, revised October 2026. https://www.nber.org/papers/w35720 (also SSRN abstract 7426447: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7426447)
  3. Crouch, Dennis. “Who Keeps the Skill: A Randomized Trial of AI Patent Drafting at Eleven Firms.” Patently-O, 9 September 2026. https://patentlyo.com/patent/2026/09/who-keeps-the-skill-a-randomized-trial-of-ai-patent-drafting-at-eleven-firms.html
  4. Lichtenberg, Nick. “The economist behind the 'China shock' tackles Gen Z's AI paradox.” Fortune, 10 October 2026. https://fortune.com/2026/10/10/david-autor-china-shock-nber-google-paper-illusion-of-competence-gen-z/
  5. PwC. “AI reshapes global labour market into two distinct paths, rewarding human skills: PwC 2026 Global AI Jobs Barometer.” Press release, London, 15 June 2026. https://pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-jobs-barometer.html (syndicated copy: https://kyodonewsprwire.jp/release/202606150856)
  6. PwC. “2026 Global AI Jobs Barometer: Global Findings.” June 2026. https://www.pwc.com/gx/en/issues/artificial-intelligence/job-barometer/2026/2026-global-ai-jobs-barometer-global-findings.pdf
  7. PwC Belgium. “AI reshapes labour market and urges for upskilling: PwC's AI Jobs Barometer 2026.” Press release, Diegem, 15 June 2026. https://press.pwc.be/ai-reshapes-labour-market-and-urges-for-upskilling-pwcs-ai-jobs-barometer-2026
  8. Katanich, Doloresz. “Human skills increasingly in demand as AI reshapes labour market, PwC finds.” Euronews, 16 June 2026. https://euronews.com/business/2026/06/16/human-skills-increasingly-in-demand-as-ai-reshapes-labour-market-pwc-finds
  9. Let's Data Science (LDS Team). “PwC Finds AI Creates Two-Track Global Labour Market.” 15 June 2026. https://letsdatascience.com/news/pwc-finds-ai-creates-two-track-global-labour-market-550ec8b1
  10. Mincer, Jacob. Schooling, Experience, and Earnings. New York: National Bureau of Economic Research / Columbia University Press, 1974. Chapter 5, “The Human Capital Earnings Function”, pp. 83-96. https://ideas.repec.org/b/nbr/nberbk/minc74-1.html ; https://www.nber.org/chapters/c1767.pdf
  11. Felten, Edward; Raj, Manav; Seamans, Robert. “Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses.” Strategic Management Journal, 42(12), 2195-2217, 2021. DOI: 10.1002/smj.3286. https://econpapers.repec.org/RePEc:bla:stratm:v:42:y:2021:i:12:p:2195-2217
  12. Eloundou, Tyna; Manning, Sam; Mishkin, Pamela; Rock, Daniel. “GPTs are GPTs: Labor market impact potential of LLMs.” Science, 384(6702), 1306-1308, June 2024. DOI: 10.1126/science.adj0998. Preprint: https://arxiv.org/abs/2303.10130
  13. Polanyi, Michael. The Tacit Dimension. Garden City, NY: Doubleday, 1966 (reissued University of Chicago Press, 2009).
  14. Autor, David H. “Polanyi's Paradox and the Shape of Employment Growth.” NBER Working Paper No. 20485, September 2014. DOI: 10.3386/w20485. https://www.nber.org/papers/w20485
  15. Brynjolfsson, Erik; Chandar, Bharat; Chen, Ruyu. “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence.” Stanford Digital Economy Lab, first released August 2025, revised 12 August 2026. https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/
  16. Federal Reserve Bank of New York. “The Labor Market for Recent College Graduates.” Data for 2026:Q3. https://www.newyorkfed.org/research/college-labor-market
  17. Brynjolfsson, Erik; Li, Danielle; Raymond, Lindsey. “Generative AI at Work.” Quarterly Journal of Economics, 140(2), 889-942, 2025. Earlier version: NBER Working Paper No. 31161, 2023. https://www.nber.org/papers/w31161 ; https://econpapers.repec.org/RePEc:oup:qjecon:v:140:y:2025:i:2:p:889-942
  18. Bastani, Hamsa; Bastani, Osbert; Sungu, Alp; Ge, Haosen; Kabakcı, Özge; Mariman, Rei. “Generative AI without guardrails can harm learning: Evidence from high school mathematics.” Proceedings of the National Academy of Sciences, 122(26), e2422633122, 2025. DOI: 10.1073/pnas.2422633122. https://econpapers.repec.org/RePEc:nas:journl:v:122:y:2025:p:e2422633122
  19. Deming, David J.; Noray, Kadeem. “Earnings Dynamics, Changing Job Skills, and STEM Careers.” Quarterly Journal of Economics, 135(4), 1965-2005, 2020. DOI: 10.1093/qje/qjaa021. https://www.hks.harvard.edu/publications/earnings-dynamics-changing-job-skills-and-stem-careers
  20. US Department of Education, Federal Student Aid. “(GENERAL-26-57) Federal Student Aid Posts Updated Reports to FSA Data Center.” Electronic announcement, 22 September 2026. https://fsapartners.ed.gov/knowledge-center/library/electronic-announcements/2026-09-22/federal-student-aid-posts-updated-reports-fsa-data-center
  21. Student Loans Company. “Student Loans in England: Financial Year 2025-26.” Official statistics, GOV.UK, updated 2 July 2026. https://www.gov.uk/government/statistics/student-loans-in-england-2025-to-2026/student-loans-in-england-financial-year-2025-26
  22. House of Commons Library. “Student loan statistics.” Research briefing SN01079, 2026. https://commonslibrary.parliament.uk/research-briefings/sn01079

Tim Green

Tim Green UK-based Systems Theorist & Independent Technology Writer

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

ORCID: 0009-0002-0156-9795 Email: tim@smarterarticles.co.uk

Listen to the free weekly SmarterArticles Podcast

Support What We Do

Discuss...