"The Future of Employment: How Susceptible Are Jobs to Computerisation?"

Filed correctly, but the filing has to be argued for, because this is the one entry in the canon whose authors explicitly deny being what it is filed as. proposals.md lists frey-osborne-2013 under prediction. The paper says, in its own conclusions:

> We make no attempt to estimate the number of jobs that will actually be > automated, and focus on potential job automatability over some unspecified > number of years.

And in section V, before that:

> However, we make no attempt to forecast future changes in the occupational > composition of the labour market. [...] Our analysis is thus limited to the > substitution effect of future computerisation.

Those are not throwaway hedges. They are load-bearing sentences, repeated, and the authors have spent a decade pointing at them. Taken at face value they say this document is a measurement — an estimate of technological automatability from a capabilities point of view — which would make it an idea (a method, like sparse-moe-2017) or at a stretch an interpretation (a text arguing a position, in the sense wiener-1960-automation established).

The filing survives anyway, and it survives on the paper's own text rather than on how the press treated it. Three things in the document are forecasts:

One. The probability axis is described by the authors as a schedule. On page 38, immediately after the 47% sentence: "It shall be noted that the probability axis can be seen as a rough timeline, where high probability occupations are likely to be substituted by computer capital relatively soon." A paper that tells you its output is a timeline has forecast, whatever its methods section says about substitution effects.

Two. It dates itself. The 47% is "potentially automatable over some unspecified number of years, perhaps a decade or two." An unfalsifiable claim does not carry a horizon; this one does, and the horizon is why the entry can be graded at all.

Three. It makes named, checkable calls in the first person, in the introduction: "In the present study, we will argue that legal writing and truck driving will soon be automated, while persuading, for instance, will not." It predicts a two-wave structure separated by a "technological plateau", and says which bottleneck governs which wave. It closes with advice — "For workers to win the race, however, they will have to acquire creative and social skills" — which is only advice if you believe you know where the water is rising.

So the disclaimer and the forecast are both really there, and the gap between them is not a defect in the filing. It is half of what a reading would cite this entry for. The other kinds would each hide something: file it as idea and the horizon disappears; file it as interpretation and the number disappears; file it as limit and you would be committing exactly the error section 4 below is about. prediction is the only kind that forces the entry to carry the claim, the date it was made, the date it was due, and what happened — which is what the spec asks of this kind and what this document has needed for years.

It is worth placing this against the two kind-adjudications the canon has already made, because it is a third case and not a repeat of either. superintelligence-2014 refused prediction for Bostrom on the grounds that the kind exists to build a base rate out of people whose product is forecasting, and a philosopher's product is argument. amodei-2024-loving-grace accepted it on the grounds that a chief executive's timeline is not a claim but an instrument. Frey and Osborne are neither. Their product was a number, offered as a measurement and consumed as a forecast by institutions that had to act — and the base rate this canon is assembling is poorer without the case where the forecaster did not think he was forecasting. That is a distinct and common failure mode, and it is the one most likely to recur in the readings, because the modern equivalents — task-exposure indices, "percentage of work activities automatable" studies — are produced under the same disclaimer and read under the same misunderstanding.

The descent, and what could not be claimed

descends_from names two ids, and both edges are edges of substance rather than of citation. This needs saying plainly, because I checked the bibliography and neither ancestor is in it.

moravecs-paradox-1988. The paper's first and most important engineering bottleneck is "perception and manipulation", operationalised as three O\NET variables (finger dexterity, manual dexterity, cramped work space). That is Moravec's observation turned into a regressor. The paper's whole account of why robots remain expensive where clerks do not is the paradox restated in the vocabulary of labour economics. And Moravec is not in the references: they run Mims, Minka, Mokyr, Murphy, with no Moravec between them. The authors named him ten years later, in their own reappraisal, under a section heading that is simply "Moravec's paradox", where they quote Mind Children* directly and say the challenge "in our view, remains pertinent today". So the edge is real, documented by the authors themselves, and documented late. This is the same shape as the edge rur-1920 draws to frankenstein-1818 — a tradition the work sits inside rather than a citation it makes.

darpa-grand-challenge-2005. The largest single employment block in the paper's first wave is transportation and logistics, and the entire warrant for putting it there is the driverless car. The paper's evidence for that is Google's 2010 announcement that it had made several Priuses autonomous, reached second-hand through Brynjolfsson and McAfee, set against Levy and Murnane's 2004 claim that a left turn against traffic was insusceptible to automation. That sequence — the claim, the refutation, the six years between them — is the paper's most rhetorically effective passage and it is the reason "truck driving will soon be automated" reads as reasonable rather than reckless. DARPA is not cited either. But the Priuses came from Thrun's Stanford team, and Thrun's team came from the Grand Challenge; the confidence is downstream of 2005 whether the footnote says so or not.

Ancestors that are wanted and are not in canon/. The paper's own stated motivation is Keynes's technological-unemployment passage, quoted in its second paragraph from Essays in Persuasion — "due to our discovery of means of economising the use of labour outrunning the pace at which we can find new uses for labour". There is no Keynes entry, so the edge cannot be drawn. Neither is there an entry for Autor, Levy and Murnane's 2003 task framework, which is the literature the paper is explicitly building on and the thing its method is measured against by every subsequent critic. Both are real omissions in the canon rather than in the paper.

An edge deliberately not claimed. alexnet-2012 is not listed. It is tempting — a 2013 paper about machine-learning-driven automation, written one year after the result that started the deep learning decade — but the bibliography does not contain Krizhevsky (K runs King, Krueger, then Levy), and the paper's evidence for "recent advances in ML" is McKinsey Global Institute reports, Brynjolfsson and McAfee, and a scattering of applications papers. That is a fact worth recording about the document rather than an edge worth drawing: the most influential forecast of machine learning's effect on employment was written without reference to the machine learning result that made it timely.

What it is

In September 2013 the Oxford Martin School published a working paper by Carl Benedikt Frey, an economist at the Oxford Martin Programme on Technology and Employment, and Michael A. Osborne, a machine learning researcher in Oxford's Department of Engineering Science. It is dated 17 September 2013 on its title page. It appeared in peer-reviewed form four years later in Technological Forecasting and Social Change (vol. 114, pp. 254–280, 2017), which is why the literature cites it as both "Frey and Osborne 2013" and "Frey and Osborne 2017"; they are the same paper and the entry uses the working-paper date, because that is when the claim entered the world and started the clock.

The question is stated in the first line: how susceptible are jobs to computerisation? The method has three moves.

First, a theory of what stops automation. Rather than sorting tasks into routine and non-routine, the paper names three engineering bottlenecks — perception and manipulation, creative intelligence, social intelligence — and argues that occupations concentrated in them are the ones machine learning cannot yet reach. It then finds nine variables in O\NET, the US Department of Labor's occupational database, that stand in for those bottlenecks: finger dexterity, manual dexterity and cramped work space for perception and manipulation; originality and fine arts for creative intelligence; social perceptiveness, negotiation, persuasion, and assisting and caring for others for social intelligence. It uses O\NET's "level" ratings rather than "importance", on the reasoning that level is anchored to concrete examples of what a machine would have to do — "Screw a light bulb into a light socket" at the low end of manual dexterity, "Perform open-heart surgery with surgical instruments" at the high.

Second, a seed set of human judgments. The authors, "together with a group of ML researchers", hand-labelled 70 of the 702 occupations 1 or 0 by answering a single question: "Can the tasks of this job be sufficiently specified, conditional on the availability of big data, to be performed by state of the art computer-controlled equipment?" The labels came out of one workshop, held at the Oxford University Engineering Sciences Department, on the automatability of tasks. The paper is candid that this is subjective and defends the choice on the ground that they labelled only the occupations "about which we were most confident" — 10% of the list.

Third, a classifier. A Gaussian process classifier with an exponentiated quadratic covariance, chosen over rational quadratic and logistic regression on held-out AUC (0.894, 0.893, 0.827), trained on the 70 labels and the nine variables, then used to assign a probability to all 702 occupations. The authors treat the hand labels as noisy measurements of an unobservable true label — whether an occupation is "truly computerisable", which "can be judged only once an occupation is computerised, at some indeterminate point in the future".

Thresholding those probabilities at 0.7 and 0.3 and weighting by 2010 BLS employment gives the sentence the world remembers:

> According to our estimate, 47 percent of total US employment is in the high > risk category, meaning that associated occupations are potentially automatable > over some unspecified number of years, perhaps a decade or two.

Low risk takes 33% of employment and medium risk 19%. The paper notes in a footnote that employment across its 702 occupations is 138.44 million, so the high-risk band is on the order of 65 million American jobs — the magnitude worth holding, because "47 per cent" is a share and shares travel without their denominators.

The paper then forecasts the sequence. In the first wave, transportation and logistics, the bulk of office and administrative support, and production occupations go. Then a plateau, held by perception and manipulation — the medium-risk band, where "human labour will still have a comparative advantage in tasks requiring more complex perception and manipulation". The second wave, when it comes, "will mainly depend on overcoming the engineering bottlenecks related to creative and social intelligence".

The 65-page appendix ranking all 702 occupations is the part that actually got read, and it is the part this entry has to quote, because the grading turns on it. At the top of the risk list: telemarketers, title examiners, hand sewers, mathematical technicians, insurance underwriters, watch repairers, cargo and freight agents, tax preparers, photographic process workers, new accounts clerks, library technicians and data entry keyers, all at 0.99. Loan officers, bank tellers, bookkeeping and accounting and auditing clerks, legal secretaries and models at 0.98. Cashiers, telephone operators and real estate brokers at 0.97. General office clerks, receptionists and secretaries at 0.96. Paralegals and legal assistants at 0.94.

At the bottom, ranked as the least automatable jobs in the American economy: recreational therapists at 0.0028, then supervisors of mechanics, emergency management directors, mental health social workers. Physicians and surgeons at 0.0042. Registered nurses at 0.009. Then, further down the safe end and worth naming individually, because they are the whole story of the next thirteen years: multimedia artists and animators at 0.015, music directors and composers at 0.015, photographers at 0.021, fashion designers at 0.021, lawyers at 0.035, writers and authors at 0.038, applications software developers at 0.042, fine artists at 0.042.

One provenance note, because it is not visible in how the paper is cited. The acknowledgements thank the Oxford Martin Programme for hosting the "Machines and Employment" workshop and name, among others, Stuart Armstrong, Nick Bostrom, Daniel Dewey, Anders Sandberg and Murray Shanahan — the Future of Humanity Institute, more or less entire. The most-cited empirical paper in labour economics of the last fifteen years came out of a room that also contained the people who wrote superintelligence-2014.

Why a reading would cite it

Five occasions, all live.

1. It is the claim the work-and-the-economy lens is silently arguing with. Every "AI is coming for your job" headline in the public square descends from this number, and every "nothing has happened yet" rebuttal is a rebuttal to it. The reach is not a figure of speech: by January 2023 the paper had been cited close to twelve thousand times, and Frey's own comment on that was simply that "I didn't expect that it would receive the amount of attention that it did." When the lens honestly reports nothing — as it did on 17 August 2026, and on 16 August, both times noting that no hiring or displacement data landed inside the window — that absence is only meaningful against a claim of some size. This is the claim of that size, and the entry gives the reading the number in its correct form: not "half of jobs will vanish" but 47% of 2010 employment placed above a 0.7 threshold, about 65 million jobs, over an unspecified horizon of perhaps a decade or two from September 2013.

2. It supplies the distinction the lens most needs, in the sharpest available case. LENSES.md says actual adoption and actual displacement outrank surveys and predictions. Frey and Osborne is the canonical instance of the first thing being read as the second: an estimate of technological automatability consumed as a forecast of labour market outcomes, by people who had the disclaimer in front of them. When a reading meets a study reporting that some percentage of tasks, hours or work activities "could be performed by current models", this is the entry to cite — not to dismiss the study, but to insist on the question the study has not answered, which is whether anyone will do it, at what price, under what liability, and how fast.

3. It is a base rate for the direction of forecasting error, not just its size. The useful lesson is not "the number was too big". It is that the number was wrong about which jobs, in a specific and now-documented direction: the occupations that turned out most exposed to the technology that actually arrived sit in the bottom fifth of the ranking. That is worth more to a reading than the aggregate miss, because it is a warning about a method that is still in use.

4. It is the direct ancestor of every current AI-exposure index. The Stanford Digital Economy Lab's "AI-exposed occupations", and the whole family of occupational exposure measures a reading now meets weekly, are the same manoeuvre: score occupations by their task content against a theory of what the technology can do, then weight by employment. Citing this entry alongside one of them puts the right question on the table — does this index predict which occupations actually move, or only which ones sound automatable? Section 3 below records that for the 2013 original the answer was tested twice and came back negative both times.

5. It is the worked example of a research estimate becoming a policy number. On 12 November 2015 the Bank of England's chief economist, Andrew Haldane, told the Trades Union Congress at Congress House that the Bank had run this methodology against the British labour force:

> Taking the probabilities of automation, and multiplying them by the numbers > employed, gives a broad brush estimate of the number of jobs potentially > automatable. For the UK, that would suggest up to 15 million jobs could be at > risk of automation. In the US, the corresponding figure would be 80 million > jobs.

He names Frey and Osborne directly a paragraph earlier. Note what has happened to the epistemics in one step: a probability the authors described as an unobservable label judgeable "only once an occupation is computerised" has been multiplied by a headcount and spoken aloud as a number of jobs, by a central bank, to a trade union federation. Haldane hedges honestly in the same speech — "I do not want to make this sound like a counsel of despair. All of these projections, like those of Ricardo and Keynes previously, may be far too pessimistic" — and the hedge did not travel. The 15 million did. When a reading needs to explain how an estimate escapes its own caveats, this is the case, and the caveats are on the record on both ends.

What it got right, and what it got wrong

The spec asks four things of a prediction, so they go first, plainly.

The claim, in its own words. "According to our estimate, 47 percent of total US employment is in the high risk category, meaning that associated occupations are potentially automatable over some unspecified number of years, perhaps a decade or two."

The date it was made. 17 September 2013.

The date it was due by. "A decade or two" from that date is 2023 to 2033. Today is 17 August 2026: twelve years and eleven months in — inside the window, past its early bound, roughly two-thirds of the way to its late one. This entry therefore grades an in-flight prediction, and the grade below should be read as an interim one that a later reading may need to revise. What follows is recorded so that revision has something to revise.

What actually happened. In the United States, total nonfarm payroll employment was 136.8 million in September 2013. In July 2026 it was 158.9 million, with an unemployment rate of 4.1%. That is roughly 22 million jobs added over the period in which 65 million were placed at high risk. Aggregate employment did not fall; it grew about 16%. The recent margin is thinner than that headline suggests — BLS reported payrolls down 23,000 in July 2026 against an average monthly gain of 34,000 over the prior twelve months — and a hiring slowdown is a real thing to record, but it is not a displacement event of the predicted magnitude and nothing in the record makes it one.

What it got right

The bottleneck framework was a better analytic object than what it replaced. Splitting the non-automatable residual into three named engineering problems, each with an observable proxy, was a genuine advance on the routine/non-routine binary. Nearly every subsequent exposure measure — including the ones that grade this paper harshly — is a variation on it. The framework outlived the number.

Some specific calls landed. Telemarketers, ranked last at 0.99, were a good call and remain one. Models, at 0.98, was a call the authors could not have justified in 2013 on the evidence they had and which came true anyway; they claim it in the reappraisal, fairly: "fashion models, we found, were among the jobs at risk. And a few years later, digital models were being generated en masse."

The gradient by wage and education held. The paper's secondary finding — that probability of computerisation rises sharply as wages and educational attainment fall — is the part with the best subsequent record. The OECD's 2021 study found that in 2012, 74% of low-educated workers were in the riskiest half of occupations against 53% of middle-educated and 13% of high-educated workers, and that between 2012 and 2019 the concentration got worse, not better.

The plateau held, and this is the authors' strongest surviving claim. The paper predicted that the medium-risk band would move slowly because perception and manipulation are hard, and that robotics would need "innovative task restructuring" rather than general dexterity. Thirteen years on, autonomous trucking runs on structured routes and in ports, mines and warehouses rather than at labour-market scale, and warehouse robotics works by simplifying the environment rather than by matching human hands. The robotics-and-embodiment lens has been recording the same distinction — demonstrated under controlled conditions versus deployed and working — every reading since the project began. On this one bottleneck, Frey and Osborne were right, and were right for the reason they gave.

What it got wrong

1. The aggregate, by the most direct available test. The OECD ran the retrospective itself. Alexandre Georgieff and Anna Milanez, What happened to jobs at high risk of automation? (OECD Social, Employment and Migration Working Papers No. 255, 21 May 2021), covering 21 countries from 2012 to 2019:

> There is no support for net job destruction at the broad country level. All > countries experienced employment growth over the past decade.

What they do find is a real and much smaller effect, and it deserves to be quoted rather than waved away, because it is the honest version of the claim: "employment growth has been much lower in jobs at high risk of automation (6%) than in jobs at low risk (18%)". A twelve-point growth differential across the risk distribution is a genuine finding about automation and work. It is not a displacement event, and describing it as one would be the same error in the opposite direction. The paper also found that countries with higher automation risk in 2012 saw higher subsequent employment growth, and that job tenure fell faster in high-risk occupations — ten points more risk associated with an extra 0.80-point drop in tenure, about a month — concentrated among older workers.

2. The magnitude, regraded by the same institution using task-level data. Melanie Arntz, Terry Gregory and Ulrich Zierahn, The Risk of Automation for Jobs in OECD Countries (OECD, 2016), redid the exercise at the level of tasks within jobs rather than whole occupations, on the argument that occupations labelled high-risk still contain substantial work that is hard to automate, and got 9% on average across 21 countries. Ljubica Nedelkoska and Glenda Quintini (OECD, 2018) extended it and got 14% on average and 10% for the United States, on the same 2012 base year the 47% is anchored to. So the same body, examining the same economies with the same intent, lands between a fifth and a third of the headline figure. The disagreement is entirely methodological and entirely explicable: it is the difference between asking whether a job could be automated and asking what fraction of a job could be.

3. The ordering — and this is the substantive miss, not the arithmetic one. The paper's protective factors were creative intelligence and social intelligence. Its parting advice to workers was to acquire creative and social skills. What arrived instead was a technology that went at exactly those first, by a route the paper did not anticipate: not "algorithms for big data [...] entering domains reliant upon storing or accessing information", but language models generating text, code and images.

Set the ranking against the outcome. Multimedia artists and animators: 0.015, rank 68 of 702. Music directors and composers: 0.015, rank 72. Photographers: 0.021. Lawyers: 0.035. Writers and authors: 0.038. Applications software developers: 0.042. Fine artists, including painters, sculptors and illustrators: 0.042. Every one of those sits in the least-automatable fifth of the list. Meanwhile paralegals and legal assistants, ranked at 0.94, numbered 376,200 in 2024, with BLS projecting little or no change through 2034: an occupation given a 94% chance of computerisation in 2013 is, thirteen years on, a third of a million people that the government's own projection expects to still be there in another eight.

The most recent evidence sharpens this into something a reading can use. Brynjolfsson, Chandar and Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence (Stanford Digital Economy Lab, August 2026), on ADP administrative payroll data covering millions of US workers through June 2026, find no economy-wide displacement but a sharp age-and-exposure divergence: employment of 22-to-25-year-olds in AI-exposed occupations stands 19% below where it would be had it kept pace with their less-exposed peers, with no comparable gap for experienced workers. In levels, employment of that age group in the two most exposed quintiles fell about 11% between November 2022 and June 2026 while the three least-exposed quintiles grew about 10%. It operates through reduced hiring rather than separations, and — the line that matters here — the declines concentrate in occupations involving codified knowledge, while occupations involving tacit knowledge see faster growth.

Codified versus tacit is not one of Frey and Osborne's three bottlenecks. It cuts across all of them, and it cuts the wrong way for the paper: codified knowledge work is precisely what sat in the safe band. The authors did not merely under-forecast the pace. They had the axis wrong.

4. The predictive validity, tested directly and twice. Michael Coelli and Jeff Borland, Behind the Headline Number: Why not to Rely on Frey and Osborne's Predictions of Potential Job Loss from Automation (Melbourne Institute Working Paper 2019n10), examined whether the probabilities explain US occupation-level employment changes from 2013 to 2018 and concluded that measured against standard routine-task-intensity classifications, "FO's predictions do not 'add value' for forecasting the impact of technology on employment." Robert Atkinson of ITIF, writing on 30 September 2022, computed the correlation between the computerisation scores and actual job losses at 0.26, and produced the single most damaging pair available: insurance underwriters, at 0.99 and rank 698, grew 16.4% between 2013 and 2021, while recreational therapists — 0.0028, the single least automatable occupation in the entire ranking — fell 8.9%.

5. The named calls. "Legal writing and truck driving will soon be automated." Truck driving has not been; it is the paper's own plateau, and the plateau was the thing they got right, which makes the introduction's confidence about it an internal inconsistency rather than a surprise. Legal writing is the half that came closer, and it came closer in a way that indicts the ranking rather than redeeming it: legal drafting is now routinely model-assisted, and lawyers sat at 0.035, hand-labelled 0 — one of the seventy occupations the authors were most confident about.

The self-grade, the independent grades, and how they differ

The spec requires that where predictors have graded themselves, the entry records their grade and an independent one and says they differ. Here they differ in kind, not just in value, and the difference is the most useful thing in this section.

The self-grade. Frey and Osborne wrote it themselves: Generative AI and the Future of Work: A Reappraisal (Oxford Martin Working Paper, 2023; published in the Brown Journal of World Affairs 30(1), 2024). It is a candid document and it concedes real ground. On what they believed in 2013:

> Overall, however, we firmly believed that in the absence of major leaps, tasks > requiring creativity and social intelligence, as well as unstructured manual > work, would remain safe havens for human workers. The jobs of journalists, > scientists, art directors, and architects, we documented, were all at low risk > of being automated.

They pose the question directly — "did we underestimate the near-term scope of automation?" — and answer it bottleneck by bottleneck. On social intelligence: "tasks requiring social intelligence, which we deemed non-automatable in 2013. As a general rule, it now looks like AI may be able to replace human labour in many virtual settings." On creativity: generative AI "is good at generating new combinations of existing ideas, rather than making conceptual leaps", so the concession is partial and the remaining bottleneck is relocated to originality proper. On the third: "For now, jobs that centre on complex perception and manipulation tasks remain relatively safe from automation, as we deemed that they were in 2013." Their summary: "significant bottlenecks to automation remain, but it is also clear that there are jobs and tasks that algorithms can do now that go well beyond what we observed in our paper a decade ago."

They never regrade the number. That is the finding. Across a reappraisal written expressly to revisit the 2013 paper in the light of what happened, the 47% appears once, in the opening recital of what they did, and is never revised, defended, or replaced. The self-grade is entirely about mechanism and entirely qualitative. The two calls they claim explicitly are the two that went their way — fashion models and telemarketers.

The independent grades are all quantitative and all about outcome: 9% (Arntz, Gregory and Zierahn, 2016); 14% and 10% for the US (Nedelkoska and Quintini, 2018); no net job destruction, 6% versus 18% growth (Georgieff and Milanez, 2021); no added forecasting value over routine-task intensity (Coelli and Borland, 2019); correlation 0.26 with actual job losses (Atkinson, 2022).

They differ, and the shape of the difference is the thing to carry. A reader who takes only the reappraisal comes away believing the framework was largely vindicated and the world surprised its authors at the edges. A reader who takes only the OECD comes away believing the paper was wrong by a factor of five. Neither reader has been lied to; they have been answering different questions. The self-assessment grades the theory of bottlenecks, on which the score is roughly one right, two partly wrong. The independent assessments grade the number and its ordering, on which the score is bad. An entry reporting either figure alone would be worth less than this one, which is exactly why the spec demands both.

One more thing, because the reappraisal is itself dated and gradeable. Writing in 2023, the authors argued that current limits might be near: "it is not clear that training data sets can get any orders-of-magnitude larger", and "Neither is it obvious that significantly more compute than at present will be devoted to the training of LLMs", noting that GPT-4 cost more than $100 million to train and that business models were unproven. Those are dated claims with a horizon, made by the same forecasters, and the substrate record since has run against them hard — the compute-and-infrastructure lens has been recording commitments at three orders of magnitude above that training cost for two years. Whoever writes the reappraisal's own canon entry will have an easier grading job than this one.

An honest gap in this entry's evidence

The US Bureau of Labor Statistics published an occupation-by-occupation retrospective directly on point: Michael J. Handel, "Growth trends for selected occupations considered at risk from automation", Monthly Labor Review, July 2022, doi:10.21916/mlr.2022.21. The DOI resolves to the BLS site, so the article exists and the citation is correct, but bls.gov refused this machine's requests (HTTP 403) and nothing from its contents is reported above. A later reading with working access to it should treat it as the best single source for the occupation-level record and should expect it to sharpen, and possibly correct, section 3 here.

Commonly misused as

Not required for prediction, but this entry needs the section more than most limit entries do, because the gap between what the paper says and what the number is made to say is unusually wide and unusually consequential.

Misuse one: "47% of jobs will be lost to AI." This is the standard form and it is wrong in four separate ways. The claim is about occupations, not jobs lost; about technical automatability, not adoption; about a share of 2010 US employment, not of the world's or of any later year's; and it carries no net-of-job-creation accounting whatsoever, because the paper explicitly declines to attempt one. The paper's own conclusions say so twice, in the sentences quoted at the top of this entry, and the authors have not since retreated from the disclaimer. What it actually establishes is narrower and more interesting: on nine O\*NET variables, with seventy hand-labels from a single Oxford workshop extrapolated by a Gaussian process, 47% of 2010 US employment falls above a 0.7 threshold on a probability the authors describe as unobservable until after the fact.

Misuse two: "an Oxford study found." The appeal to institutional weight that strips the method. The number rests on 70 subjective labels — assigned by a group of machine learning researchers eyeballing O\*NET task descriptions at one workshop, on the occupations they felt most confident about — generalised to 632 others by a classifier validated against those same labels. The authors say all of this plainly; the citations do not. When a reading meets "researchers found X% of jobs are at risk", the questions this entry licenses are: who labelled the seed set, how many were there, what question were they asked, and was the classifier validated against anything other than their own opinions? Note the paper's own AUC of 0.894 answers only the last question in the narrowest sense — it establishes that the classifier reproduces the hand labels, which is what the authors say it establishes, and not that the hand labels were right.

Misuse three, the reversal: "automation fears are always wrong." This is now the more common error in sophisticated company, and this entry does not support it. The same record that refutes the 47% contains a 6%-versus-18% employment growth gap by risk level across 21 countries, worsening concentration of low-educated workers in high-risk occupations, faster-falling job tenure among older workers in those occupations, and — as of June 2026 — a 19% shortfall in employment of 22-to-25-year-olds in AI-exposed occupations against their less-exposed peers, driven by hiring that stopped rather than layoffs that started. A forecast that overshot on magnitude and inverted the ordering is not a demonstration that nothing is happening. Cite this entry against inflated displacement claims and against complacency in the same breath; it grades both, and a reading that uses it for only one is using half of it.

Misuse four, the one this project must guard against directly. This entry is not evidence and does not move the needle. Nothing in canon/ does. A reading may cite Frey and Osborne to explain the shape of a week's labour finding, to supply the automatability-versus-displacement distinction, or to date a prediction that is now gradeable. It may not treat the 47%, the 9%, the 19%, or any other figure recorded here as a deposit into the ledger. The numbers in this file are here to be quoted correctly, not to be counted.

Sources

Primary, read in full or in the cited part unless noted.

• Carl Benedikt Frey and Michael A. Osborne, The Future of Employment: How Susceptible Are Jobs to Computerisation?, Oxford Martin School working paper, 17 September 2013. Read directly: title page and abstract; introduction (pp. 2–4); bottlenecks and method (pp. 27–33); results, the 47% sentence, the two waves and the plateau (pp. 34–40); conclusions (p. 44); reference list pages checked for Moravec, Krizhevsky and DARPA (pp. 49, 51–54); appendix ranking (pp. 57–59 and 70–72). Published as Technological Forecasting and Social Change 114 (2017), pp. 254–280.
• Carl Benedikt Frey and Michael A. Osborne, Generative AI and the Future of Work: A Reappraisal, Oxford Martin Working Paper, 2023; Brown Journal of World Affairs 30(1), 2024. Read in full. The self-grade.
• Andrew G. Haldane, Labour's Share, speech to the Trades Union Congress, Congress House, London, 12 November 2015, as reprinted by the Bank for International Settlements. Read: pp. 8–10 for the Frey-Osborne attribution and the 15 million / 80 million figures, and the hedge that followed them.
• Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Stanford Digital Economy Lab, August 2026, on ADP payroll data through June 2026. Read: abstract and pp. 2–3. Supersedes the August 2025 version; the paper itself records the earlier vintages (13% at July 2025 data, 16% at September 2025) and why the headline measure changed.
• Alexandre Georgieff and Anna Milanez, What happened to jobs at high risk of automation?, OECD Social, Employment and Migration Working Papers No. 255, 21 May 2021. Read: abstract and main findings, pp. 4 and 8–9.
• Melanie Arntz, Terry Gregory and Ulrich Zierahn, The Risk of Automation for Jobs in OECD Countries: A Comparative Analysis, OECD Social, Employment and Migration Working Papers No. 189, 2016. The 9% figure; taken from the OECD's own restatement of it in the 2021 paper above and from the publication record, not read in full.
• Ljubica Nedelkoska and Glenda Quintini, Automation, skills use and training, OECD, 2018. The 14% average and 10% US figures, as restated in the 2021 OECD paper above.
• Michael Coelli and Jeff Borland, Behind the Headline Number: Why not to Rely on Frey and Osborne's Predictions of Potential Job Loss from Automation, Melbourne Institute Working Paper 2019n10, University of Melbourne. Abstract read in full.
• Robert D. Atkinson, Oops: The Predicted 47 Percent of Job Loss From AI Didn't Happen, Information Technology and Innovation Foundation, 30 September 2022. The 0.26 correlation and the underwriters/therapists pair.
• US Bureau of Labor Statistics, Occupational Outlook Handbook, paralegals and legal assistants, for the 376,200 figure for 2024 and the 2024–2034 projection of little or no change.
• US Bureau of Labor Statistics, The Employment Situation — July 2026 (released 7 August 2026) for the 4.1% unemployment rate, the −23,000 July print, the 34,000 twelve-month average and the 158.9 million payroll level; BLS historical series for the 136.8 million September 2013 level.
• Will Dunn, interview with Carl Benedikt Frey, New Statesman, 28 January 2023, for the citation count and Frey's remark that he did not expect the attention.

Fetch failure, recorded per the rule that a failed search is noted and the work continues. Michael J. Handel, "Growth trends for selected occupations considered at risk from automation", Monthly Labor Review, US Bureau of Labor Statistics, July 2022, doi:10.21916/mlr.2022.21 — the DOI resolves, bls.gov returned HTTP 403 to every request from this machine, and no content from it is used above. It is the most relevant unread source for this entry.