AI and the future of work: what Anthropic, DeepMind and OpenAI actually claim

None of the three labs' texts says how many jobs AI will take: each one calculates something else.

In September 2026 Anthropic released "Economic Scenarios for Transformative AI" with an interactive Explorer. In its three scenarios US GDP by January 2030 is 1.6%, 8.3% and 32.4% above the no-AI path. There are no probabilities, and the authors state plainly that these are not predictions. All five authors are from the company, and there is no conflict-of-interest disclosure.

In October 2026 Google DeepMind published an essay, "The Economics of Intelligence", and a paper, "Organizing Intelligence". The authors, among them company co-founder Shane Legg, build a theory of the capability level at which cheap coordination between agents makes firms larger; there are no dates. The paper is a preprint that describes neither review nor funding, and the essay sits on a company venue, the DeepMind Institute, where Legg is managing editor. The key parameter γ has not been measured: the only estimate comes from a LessWrong post (see the section on the DeepMind model).

We put the ATLAS report on Gemini usage in the DeepMind column, but of its 18 authors 12 are from Google and 6 from Google DeepMind: it is really a Google document.

In April 2026 OpenAI published its AI jobs transition framework: the GPT-5.4 model sorts occupations into four groups and itself decides whether an occupation needs a human. The documents we read contain no GDP estimate of the company's own (one page was not accessible to us). The only macro number is in a letter to OSTP from 2025: the first trillion dollars of AI infrastructure "could" add more than 5% of GDP growth over three years, through the spending, not through productivity. There is no method, and the letter asks for tax breaks.

OpenAI's disclosures are partial. In its paper on how organisations use ChatGPT, two of the five authors are paid contractors of the company, and the paper on Codex was reviewed before publication by the chief financial officer.

A scenario cannot be tested as a forecast: it answers a "what if" question. A theory cannot be tested until its parameters are measured. A classification must not be mistaken for an effect: none of OpenAI's four documents on work estimates what AI has done to employment, wages or output.

Executives' statements are not measurements either. In March 2025 Dario Amodei (Anthropic) said that within 3–6 months AI might be writing 90% of code; Redwood Research put the share at Anthropic itself at about 50% of lines in October 2025. Sundar Pichai (Google) reports that the share of new code generated by AI rose from more than a quarter in October 2024 to 75% in April 2026, without defining new code. Ronnie Chatterji, OpenAI's chief economist, wrote in October 2025 that GPT-5-class models match or beat professionals on about half of GDPval tasks, but in OpenAI's own paper GPT-5 has 39.0% outright wins, and even the leader, a competitor's model, Claude Opus 4.1, has wins plus ties below half, at 47.6%.010101I wrote about Anthropic's chart on what share of its internal research and engineering work Claude leads, according to the company. The share of code written by AI is an even slipperier metric: every company uses its own denominator.Knizhny kub · AI accelerating its own development01I wrote about Anthropic's chart on what share of its internal research and engineering work Claude leads, according to the company. The share of code written by AI is an even slipperier metric: every company uses its own denominator.Knizhny kub · AI accelerating its own development

According to US Bureau of Labor Statistics (BLS) data, US unemployment was 4.1% of the labour force in October 2026. The Congressional Budget Office (CBO) never mentions AI in the labour-market section of its outlook for 2026–2036, but that is a property of the forecast, not a measurement: events after December 3, 2025 are not in it.

In its March paper on a new exposure measure, Anthropic used the CPS monthly household survey to compare unemployment in occupations with heavier Claude use against the rest. By the company's own estimate, unemployment doubling from 3% to 6% in exposed occupations would show, and the detection threshold is about 1 percentage point relative to the rest; the authors call the threshold rough. The data show no gap of 1 percentage point or more. So "no effect found" means "no effect larger than the threshold".

The most rigorous result is Danish. Humlum and Vestergaard linked a survey of 25,000 workers at 7,000 workplaces in 11 exposed occupations to administrative registers through December 2024. Confidence intervals rule out chatbot effects on earnings and hours above 2%. The limitation: this predates agents.

The other measurements are cruder; their caveats are in Table 2. The OECD and the Yale Budget Lab tracker see no clear AI footprint yet, in youth unemployment and in employment respectively, but the OECD's check has no regression and the OECD itself calls its conclusion preliminary, while Yale gives no values at all. The one estimate pointing the other way is a St. Louis Fed model decomposition that assigns AI about a third of the rise in unemployment among 18–24-year-olds from April 2023 to December 2025; but AI is measured there by skills in job postings, and the source was not read in full.

Firms themselves report few AI-driven cuts. Challenger counted 121,401 announced job cuts citing AI in January–October 2026, about 0.07% of 170.3 million jobs (our calculation from Challenger and BLS data); these are announcements, not actual layoffs. In the Firm Data on AI survey, 87% of roughly 9,000 executives in 7 countries see no effect of AI on employment over the past three years; this is a self-report.

Overall employment figures have not moved yet, but most series end before agents went into mass use in 2026, as the diagram below shows. What this means for Anthropic's scenarios is covered in their section.

The Stanford paper "Canaries in the Coal Mine" (August 2026) used ADP payroll data to compare 22–25-year-olds in occupations where AI can perform a large share of tasks with less exposed peers. From November 2022 to June 2026 the first group's employment fell 11% and the second's grew 10%, leaving the first 19% below its peers' trajectory. About half of the gap comes from growth in the comparison group, the ADP sample is not representative, and the authors themselves call the result descriptive. The mechanism is hiring, not layoffs: exits among exposed young workers fell at least as much as elsewhere.

The comparable gap measure widened: 15% on data through July 2025, 19% through June 2026. But in the same edition the coefficient for the gap between the extreme exposure quintiles is −0.179 without controls, and with interest rates, graduate share and remote work controlled together it is −0.080 with a standard error of 0.057, not significant. Its 95% interval runs from about −0.19 to +0.03 (our calculation from the published standard error), so it includes both zero and the original estimate.

Other data repeat the sign but not the scale (Table 3). On the ACS population survey the Canaries authors find a young-worker gap of −2.2 percentage points for 2022–2024, with an interval including zero, against −13.2 points in ADP over the same window. Firm comparisons diverge: in Hosseini and Lichtinger, junior employment at firms that adopted generative AI is 9% lower six quarters later, but adopters differed even before adopting, while in Danish registries Humlum and Vestergaard find the share of 22–25-year-olds at employers encouraging chatbots moves by no more than a third of a percentage point (data through December 2024).

Part of the decline began before ChatGPT. In data from Google economists Iscenko and Curto Millet, postings in the most exposed quintile of occupations peaked in March–April 2022; the authors work for a company that sells AI. Unemployment insurance claims show risk in exposed occupations rising from early 2022; once the Fed rate is controlled for, the post-ChatGPT variables add almost nothing to the explanation, but these data cannot see newcomers with no work history.

In sum, less hiring of young workers in exposed occupations recurs in payroll and administrative data, résumés and postings, but its size depends on source and controls, and it has more than one cause. It is the strongest of the weak signals.

The share of firms using AI depends on the question. According to the Census Bureau, 19.8% of US firms used AI to produce goods or services in May 2026. But after the wording changed on November 17, 2025, the share rose from roughly 10% to 20%, and, as the Minneapolis Fed shows, the question created that jump, not adoption. According to the St. Louis Fed, in August 2025, 37.4% of adults aged 18–64 used generative AI at work, but it accounted for 5.7% of work hours. In a survey of almost 9,000 executives in seven countries (May–July 2026), 78% of firms use AI, but 83% (on average across the seven surveys) see no effect on productivity.020202I reviewed McKinsey's 2026 report: 80% of people who use AI themselves say their productivity grew, but only 37% see any contribution to EBIT, and these are self-assessments. Same picture as in the firm surveys: access to a tool is not yet a result.Knizhny kub · McKinsey State of AI 202602I reviewed McKinsey's 2026 report: 80% of people who use AI themselves say their productivity grew, but only 37% see any contribution to EBIT, and these are self-assessments. Same picture as in the firm surveys: access to a tool is not yet a result.Knizhny kub · McKinsey State of AI 2026

Productivity does not show it yet. According to BLS, US nonfarm business productivity has grown 2.2% a year since the fourth quarter of 2019, the same pace as the long-run average since 1947. The estimate for the third quarter of 2026 is preliminary (Table 6), and one quarter proves nothing. The reading also depends on the baseline: the acceleration came in 2023–2024, before broad adoption.

Investment is where AI does show. In a FEDS Note, Fed staff added up the AI-related categories of the national accounts: software, data centers, power facilities and computers. In the first quarter of 2026 they contributed 1.18 percentage points to the annualized US GDP growth rate, or 0.73 points net of computer net exports. With GDP growth that quarter at 2.1% annualized (BEA's third estimate), that is about 35% of growth (our calculation from the authors' formulas; data before the annual revision). The accounts have no "AI" line, and this calculation, like the St. Louis Fed's in Table 6, measures demand, construction and equipment, not productivity.

In its preliminary S-1 prospectus of October 26, 2026, Anthropic disclosed minimum commitments to cloud providers, payable regardless of actual use: at least $44.9 billion to Microsoft, about $114.5 billion to Google and at least $110 billion to Amazon, over terms running to 2033–2036. This is the company's own disclosure, not spending and not capacity utilization. OpenAI has no prospectus, so there is nothing to compare.

The developers of frontier models also publish data on how those models are used: Anthropic has the Economic Index, Google has ATLAS, and OpenAI has Signals and its papers on ChatGPT usage. All of them show what people ask the model for, not what came of it. The authors work for the vendor or are its paid contractors.

Anthropic maps conversations with Claude onto tasks from the O*NET occupational database. Its main measure, m, is the share of working time in tasks where Claude is meaningfully used. In the company's March 2026 paper, on August and November 2025 data, average m is 0.14: 0.22 for cognitive occupations and 0.01 for the rest. In the August 2025 data, a mere change of labeller moves the "automation" share (Table 7).

In its January 2026 report the company estimates the contribution to productivity growth with Hulten's formula: time saved, weighted by the task's share of the economy. The savings are judged by Claude itself. The result is +1.8 percentage points a year over ten years, and after adjusting for failures (success is also judged by Claude), 1.2 for Claude.ai and 1.0 for the API. It is the vendor's model calculation, assuming today's models and usage are fully adopted, not a measurement or a dated forecast.

ATLAS covers 14.65 million interactions from April 6–19, 2026, only in Gemini's free products, without the paid API and enterprise customers, where automation is likelier. So the authors' conclusion, that there is no evidence of imminent mass white-collar automation, does not prove its absence. And on synthetic conversations generated by the same Gemini, the labeller identifies the correct O*NET task in only 22.6% of cases.

OpenAI's paper on organisations reports that monthly output tokens of enterprise customers in its panel grew about sevenfold from June 2025 to March 2026, and roughly fourfold for the same customers. But a token is neither a task nor value.

In the jobs transition framework, actual ChatGPT usage in the "high risk" group is 6.3 times that in the "fewer changes" group (our calculation from OpenAI's figure). But the high-risk group's theoretical exposure is many times higher too; adjusting for it leaves a gap of about 1.24 times (also our calculation).

The indices share one weakness: a single product, an in-house labeller, no view of the outcome. In Work at the Frontier, OpenAI itself lists five things it cannot observe, such as whether an answer was used.

The indices are the best available data on how AI is used at work, but not on its effect on employment or productivity.030303I wrote about Diogo Almeida's talk: the industry has learned to build excellent assistants, but dependable automation needs different properties. That is why splitting conversations into "automation" and "augmentation" is useful but does not tell you whether the work can be handed over.Knizhny kub · assistance versus automation03I wrote about Diogo Almeida's talk: the industry has learned to build excellent assistants, but dependable automation needs different properties. That is why splitting conversations into "automation" and "augmentation" is useful but does not tell you whether the work can be handed over.Knizhny kub · assistance versus automation

Developer productivity is not code volume but what reaches users. In an NBER working paper, Demirer, Musolff and Yang used Microsoft telemetry to follow more than 500,000 open-source developers on GitHub. Those who moved to asynchronous agents, which are handed a whole task, made 240% more commits than before AI tools but shipped only 30% more releases. Since early 2025 about twice as many new apps appear each month, but their total usage has not grown on any of four consumer marketplaces. Caveats: open source, a release is a crude measure of value, no peer review.

In corporate code, Chen and Stratton used Jellyfish platform data to compare 655 firms that adopted agents at different times. Twelve months on, beyond the effect of assistants, commits rose 36%, while the time from submitting a pull request to approval grew 56.1%, almost two days: review has become a queue. Caveats: a preprint dated November 7, 2026, not peer-reviewed, with numbers that changed between versions; only Jellyfish clients; and by July 2026, 98.3% of the firms had adopted agents, leaving almost nobody to compare against.

In a Google experiment in 2024, before agents, a time saving of about 21% is not significant after adjustments; in METR's repeat experiment the effect is not significant either (Table 8). In GDPval, OpenAI calculated that a naive "90 times faster than an expert" shrinks to 1.12–1.39 times once a human checks the result.

Anthropic reports that in August 2026, in interactive sessions started by employees, agents logged 6 times as many hours as people. This is time, not outcome: for about 80% of a session the agent waits for a human. According to the company, Claude wrote over 80% of production lines in May 2026 and over 90% of code "contributions" per the S-1 prospectus, which leaves the word undefined.

Only Anthropic let an outside evaluator into its internal data: METR estimated that capability growth has sped up 1.9 times since October 2025 (interval 1.5–2.8), while employees' self-reports said about 12 times. This is research acceleration, not product output, and Anthropic redacted most of the most informative data in the publication. Google claims speed-ups in individual teams without a baseline, period or method; OpenAI's paper on Codex shows token spend growing, not a measured outcome.

Output grows several times faster than outcome. In our assessment, acceptance now sets throughput — the review and checks after which an edit reaches the product (longread on infrastructure for agentic development) — and the gap should be closed by speeding them up, not by generating more.

By Indeed's labelling, in May 2026 US senior postings were up 14.7% year on year and entry-level postings down 7.5%. These are ads on one site, not hires, and AI is not separated from interest rates and the business cycle.

In its October report on employment, Anthropic says that since early 2025 the junior share in its technical units fell from 27% to 9% of hires and from 21% to 14% of staff. These are shares, not headcounts, "junior" is not defined, and this is one fast-growing firm.040404I reviewed Anthropic's October report on how employment is changing inside the company. Agents speed up execution, and then processes, roles and decision rules change — junior hiring is just one consequence.Knizhny kub · employment inside Anthropic04I reviewed Anthropic's October report on how employment is changing inside the company. Agents speed up execution, and then processes, roles and decision rules change — junior hiring is just one consequence.Knizhny kub · employment inside Anthropic

In an experiment by Shen and Tamkin of Anthropic, 52 developers spent one session learning an unfamiliar Python library, half with a chat assistant. On a test without AI the assisted group scored 4.15 points lower out of 27. The company calls them mostly junior, yet 29 of the 52 had seven or more years of experience: they were new only to the library.

Liu and co-authors found a similar pattern with fraction and reading problems, and Wiles and co-authors with BCG consultants on tasks from an OpenAI team: without the assistant, participants did worse than control or no better (Table 9). The longest intervention in these studies is about three hours: these test the first hours of an unfamiliar task, not a career.

How many engineers are needed depends on elasticity: how much more software is bought when it costs less. In Chen and Stratton's model, engineering employment grows only if elasticity is above 3.68. At the literature's average of 2.27, employment falls 13.2%. This is a scenario, not a forecast, and the estimate comes from studies of 2005–2014. OpenAI's US framework gives developers an elasticity of about 1.4, by a GPT model's judgment. The methods and objects differ, and none of the studies we read measures elasticity for agentic development.

In a review published by the Anthropic Institute, Roodman and Massenkoff pooled 56 randomised trials in the US: subsidised training for the unemployed adds on average 2 to 3 percentage points of employment at about $13,000 per training slot. The studies we read have no data on corporate retraining of engineers.

The site author's position on juniors is conditional ("The Junior Engineer After Code"): hire them if the company is willing to pay for reproducing expertise and builds learning as a production system; if you need a cheap senior engineer with an agent, don't. It is a rule for one firm, not a market forecast, and opposite views were voiced on the show. Two unmeasured numbers would test it: the cost of reproducing expertise and the productivity of a senior engineer paired with an agent.

Anthropic's scenarios are a "what if" calculation. The authors' model splits the US economy into tasks with three inputs: the share of tasks AI is able to do (m), the share of firms that have adopted it (d) and the gain per task (a). In mid-2026 m = 0.14, from Anthropic's own study (a measure of Claude use, not of capability), while d = 0.10 was set by the authors themselves, guided by the Census BTOS business survey.

GDP and wages for January 2030 are measured against a path without AI, a baseline nobody can observe. GDP ends up 1.6 / 8.3 / 32.4% above it in the modest, substantial and extreme scenarios. Total unemployment is 3.9 / 4.6 / 11.9%, against the model's starting level of 3.8%; nobody leaves the labour force in the model.

On average over 3.5 years, the scenarios add 0.45 / 2.3 / 8.3 percentage points a year to GDP growth (our calculation from the authors' formulas); in the final year of the extreme path growth reaches 15.4%, against 2.0% without AI. The fall of more than 10% in cognitive wages on the Explorer page is also measured against the no-AI path, not against today's pay.

Within a single scenario, two unmeasured parameters move the result several-fold. In the extreme scenario, cognitive unemployment is 2.6% when wages adjust at once (ξ = 0) and 24.0% when they barely fall (ξ = 0.9): the cost of the shock shifts from wages to jobs. GDP in the same scenario is +21.3% at a capital-supply elasticity ε = 1 and +43.3% at infinite elasticity. The baseline ξ = 0.5 and ε = 3 were chosen by the authors, who themselves write that ε has not been measured and there is no agreed value for ξ. The paper comes with no code.

A convenient measure for comparing the scenarios is m·d, the share of tasks that AI is able to do in firms that have adopted it: 1.4% for all of them in mid-2026, 1.6 / 2.0 / 2.6% in January 2027 and 2.2 / 3.9 / 7.5% in January 2028 (our calculation from the authors' formulas). The 2027 gap is probably smaller than the error of the approximate measures, although that error has not been estimated, and on unemployment our replica of the model separates only the extreme path even in 2028 (Table 11). Hence our hypothesis, not the authors' conclusion: 2027 will most likely not tell the scenarios apart.

The authors themselves acknowledge that flat employment is consistent both with the modest path and with the start of the extreme one, and they give no thresholds. That is why the last section watches the model's inputs rather than unemployment.

In "The AI Expectations Gap" survey, the answers of Anthropic's employees, run through the model, imply GDP growth of 17.1% in 2030, above the extreme scenario's 15.4%. These are beliefs, not a calibration: the unweighted answers of 1,334 employees whose earnings are tied to the product; the authors call them personal views, not a company forecast.050505When I summarised these scenarios in a post, I ran the Explorer myself, and my settings produced a result more extreme than the authors' own extreme scenario. That was about the future I would like to see, not a forecast, so it does not contradict this episode's view that change is gradual.Knizhny kub · Anthropic's three scenarios05When I summarised these scenarios in a post, I ran the Explorer myself, and my settings produced a result more extreme than the authors' own extreme scenario. That was about the future I would like to see, not a forecast, so it does not contradict this episode's view that change is gradual.Knizhny kub · Anthropic's three scenarios

What happens to firms if the workers are copies of one agent? In a paper by Imas, Koh and Legg of Google DeepMind (a preprint, October 2026), coordination is reduced to a single number γ, which describes how a team's output grows with its size. Below the threshold γ = 1, firm growth has a ceiling; above it, average output per member grows without bound and in the limit the market consolidates into one firm.

The numbers in the paper are the authors' illustrations, not estimates (see Table 12). In one simulation, for instance, 147 firms shrink to one while all 100 people stay employed. The model has no market power and no labour share.

The paper contains 0 estimated parameters and 0 dates: the “when” in DeepMind's essay means “at what capability level”, not “in what year”. The only number for γ among agents, 0.48–0.68, comes from Toby Ord's LessWrong post: a re-analysis of OpenAI's charts, not a measurement. The paper's authors only propose a test: a benchmark plus four kinds of data on firms.

What do agent teams show? In CooperBench, two agents each build their own feature in the same repository. With the best 2025 models (GPT-5 and Claude Sonnet 4.5) the pair succeeds in about 25% of cases, roughly half as often as one agent doing both tasks. There is no human baseline. In Kim and co-authors' study, multi-agent systems average −0.3% against a single agent, ranging from −70% to +80.8% by task. The DeepMind paper cites only the failures.

The outcome also depends on the scaffolding, the code around the model. In Worse Together (an Anthropic Fellows project), a GPT-5.6-sol team with a messaging channel succeeds in 56.9% of cases in one environment, against 92.6% for a single coordinator. In the same environment, three procedural instructions to the agents lift a Claude Opus 5 team to 95.3%, above the coordinator's 84.3%.

And human firms? Kim (INSEAD) and Koning (Harvard Business School) compared startups founded in 2020–2024: AI-native ones are about a quarter smaller in headcount at a comparable valuation. The comparison is cross-sectional, and the authors warn that it describes the shape of a firm, not the number of jobs. In a San Francisco Fed model (which its authors call an incomplete draft), concentration first rises and then falls as AI spreads, driven by the fixed cost of adoption, not by coordination.

DeepMind's model is a careful theorem about a mechanism, not a forecast: a forecast would need a measured γ and data on coordination inside firms. In the experiments, the cost of coordinating agents depends on scaffolding and instructions; there is no single number for it.

How much will AI add to US total factor productivity (TFP) growth? In an NBER working paper from May 2024, Acemoglu got about 0.07 percentage points a year over ten years from 2023 tasks — an upper bound that ignores new tasks and better models. Aghion and Bunel used the same formula with their own inputs and got a median of 0.68 percentage points a year. By our calculation from the authors' formulas, the 9.7-fold gap comes from three assumptions, not data: the share of exposed tasks, the share worth automating and the cost saving per task.

The second lever is the elasticity of substitution σ: how well machine tasks substitute for human ones. In a Journal of Economic Perspectives review in 2026, Jones derived the ceiling on income if half of all tasks become infinitely productive: a limit, not a path. By our calculation from his formula, income rises 1.19-fold at σ = 0.2, 2-fold at σ = 0.5 and 16-fold at σ = 0.8. Nobody has measured σ for AI across the economy. Anthropic's model uses 0.5, Jones and Tonetti's model uses 0.2, though the models also differ in structure.

Budget forecasts sit near the low end: CBO built in +0.1 percentage points a year to nonfarm business TFP for 2026–2036, and Penn Wharton expects the contribution to peak at about 0.2 percentage points in the early 2030s. The OECD estimates and the IMF scenario are in Table 13.

A survey does not narrow the spread: in Karger and co-authors' survey, 69 economists put on average a 14% chance (median 10%) on the fast scenario becoming an expert panel's most common choice at the end of 2030. That is a question about the panel's choice, not the scenario itself; about 3% of those invited responded.

Markets send a mixed signal. Expected explosive growth should push real rates up: according to Chow, Halperin and Mazlish, across 59 countries growth expectations 1 percentage point higher go with real rates roughly 1 point higher, but that is a correlation, and it covers growth expectations of any kind, not AI ones. US real rates have risen since 2021, but the rise coincided with the oil shock, deficits and policy tightening (Table 17), so it cannot be pinned on AI.

The high estimates partly come from one network of authors: Korinek co-authored the scenarios for the transition to AGI, the review with Trammell and Anthropic's model, Jones co-authored that model too, and Halperin worked on the interest-rate paper and the Karger survey. This does not prove bias, and the low estimates are not neutral either: Acemoglu has long been sceptical of automation.

The BLS measures the share of nonfarm business output that goes to workers. In its release of November 5, 2026, it was 53.0% in the third quarter of 2026, the lowest since 1947. The decline is old: in the BLS series the share was 63.8% in the first quarter of 2000 and 56.3% in the fourth quarter of 2019. A level revision and the BEA series for the whole economy are in Table 14. None of the sources we read attributes the decline to AI.

In Anthropic's scenarios the labour share falls by January 2030 to 59.4%, 56.1% and 45.2% in the modest, substantial and extreme scenarios. But they start from a model base of 60%, not the 53.0% measured by the BLS, a gap of about 7 percentage points (our calculation from the authors' and BLS figures), so only changes within each series are comparable, not levels. By the same date capital income is 3.1 / 18.9 / 81.4% above the no-AI path, labour income only 0.6 / 1.4 / 0.5%.

The other frameworks do not forecast the labour share. In a draft NBER paper (July 2025), Pascual Restrepo proves that if the set of tasks is fixed, any job needs a finite amount of compute and compute grows without limit, then the labour share tends to zero, while wages converge to a positive level and stop growing with the economy. The parameters of that limit are unmeasured. The DeepMind model has no labour share. OpenAI's programme has no number on it, although the company names the risk of gains concentrating in firms like itself.

Capital income is not everyone's income. According to the Federal Reserve's distributional accounts, in the second quarter of 2026 the top 1% of households held 50.9% of corporate equities and mutual fund shares, while the bottom half held 2.3% of all net worth. With ownership as it is, a shift towards capital is a shift towards the top.

That leaves the buyer. Erik Brynjolfsson and co-authors surveyed 1,908 Americans and put US consumer surplus from chatbots at $172.3 billion a year (March 2026), based on stated preferences rather than spending. We found no estimate of how the gains are split between vendors and business customers: the FTC press release on its report on cloud–lab partnerships describes the mechanisms without numbers (we did not read the report itself).

In its June 2026 economic policy framework, Anthropic sorts US measures into three tiers by unemployment: about 5%, about 10% and above historical peaks. For the top tier the company only lists candidates, up to levies on tokens, with no rates or amounts. It advises watching wages and the labour share too, but sets thresholds on unemployment alone, a lagging indicator.

OpenAI's April memo Industrial Policy for the Intelligence Age makes 20 proposals with no amounts or thresholds for public measures and no references to empirical work. A safety net for workers would switch on at "pre-defined thresholds", but their values are not given. The DeepMind Institute essay and the paper on firm boundaries contain no measures; a separate institute essay proposes sets of measures for different scenarios (source not read in full).

The labs' interests and caveats differ. In its S-1 prospectus, Anthropic counts taxes on AI companies among the risks to its revenue, so token levies are being floated by the company that would pay them. OpenAI's letter to OSTP asks for tax incentives. The DeepMind essays state that they are not Google's official position.

The measures in the studies we read were tested where there was demand for labour. RTAA wage insurance, in the estimate by Hyman, Kovak and Leive, raised employment among trade-displaced workers around age 50 by 8–17 percentage points in the first two years, but the employment gain was gone by year four. Guaranteed income of $1,000 a month for three years, in the experiment by Vivalt and co-authors, cut recipients' labour force participation by 4.2 percentage points. Sectoral training and other measures are in Table 15.

The 2026 laws deal with procedure, not money. From July 1, 2027, California's SB 947 bars firing workers by relying solely on an automated system. New York's layoff notice form is in Table 15.

Comparing the tiers with Anthropic's scenarios is our own step. By January 2030, unemployment in the modest and substantial scenarios is 3.9% and 4.6% (the first tier), and 11.9% in the extreme one (the second). But with fully flexible wages the extreme scenario yields total unemployment of only 3.1%, with cognitive wages 42.2% below the no-AI path. An edge case, but the unemployment threshold stays silent when the shock lands on wages.

Russia is covered neither by the WEF employer survey for 2025 (55 economies), nor by the OpenAI report (US users only), nor by the OECD Employment Outlook 2026. Russian sources measure occupational exposure and AI adoption, but not the effect of AI on employment.

How exposed are Russian occupations to AI? In a May 2026 working paper, Rostislav Kapelyushnikov (HSE University) mapped ILO task indices onto the microdata of Rosstat's 2019 labour force survey. Each occupation scores from 0 to 1; the average across all employed people is 0.30. High exposure covers 12.0% of the employed and significant exposure 31.3%; significant includes high, so the two do not add up. In information and communication the average is 0.44, and 48.9% of workers are highly exposed. Agents are not in the indices.

How many firms already use AI? In May 2026 the Bank of Russia asked about it for the first time in its enterprise monitoring survey: 14,700 non-financial companies, generative AI only, answered by management. Generative AI is used by 14% of firms; that is a share of firms, not of workers. Active use is reported by 40% of users, or about 6% of all firms (our calculation: 14% × 0.4). By their own assessment, 35% of users see higher productivity. The October survey says nothing about AI.

What about IT hiring? The hh.ru review for September 2026 shows IT vacancies down 33% year on year. But vacancies fell 32% in marketing, 31% in finance and 30% in HR. That fits both AI and a general squeeze on office work, and nobody has broken the decline down by cause. Salaries from Habr Career are in Table 16.

The state sets targets rather than measuring. The National Strategy for the Development of Artificial Intelligence sets a target of at least 80% of workers with AI skills by 2030 but does not define a skill. The law on supporting the development of artificial intelligence technologies (No. 243-FZ, in force since September 1, 2026) contains no provisions on labour or employment (source not read in full: we read the text in a legal reference system and did not check the official publication or the status of promised amendments).

The authors of Anthropic's scenarios name four quantities on which the paths should separate within a year or two: the shares of tasks and firms using AI, the gain per task and job-search time. They give no thresholds and no probabilities. DeepMind names neither dates nor thresholds, and OpenAI gives no thresholds either. So we set them ourselves.

Our forecast puts the thresholds at the midpoints between the scenario trajectories for January 2028: for the share of tasks AI can do (m) about 0.18 and 0.25, for the share of firms (d) about 0.165 and 0.22, and for m·d about 3% and 5.7% (our calculation from the authors' formulas). For d, this means the share of firms using AI in the Census BTOS survey growing from its starting 18% by roughly 1.65 and 2.2 times under unchanged question wording (our calculation, assuming d grows in proportion to the BTOS share). m can only be tracked through Anthropic's exposure measure, which has not been updated since March 2026.

For the aggregates our forecast is cruder. US nonfarm business productivity growing faster than 3% year over year for three years in a row, with a rising capital share, would separate the substantial path from the modest one; preliminary BLS data put it at 2.1% year over year in the third quarter of 2026. The extreme path needs US GDP growth above 6–8% in 2028–2029. By our calculation, unemployment will not tell the modest and substantial paths apart by 2028 (Table 11).

The most sensitive gauge is young-worker hiring, but with all controls its estimate is not significant either (see the section on young workers). In software, the gauges are closed tasks against code volume (Table 17) and review time.

The METR task-horizon series has not been updated since May 2026, and real interest rates are ambiguous (Table 17), so we set no thresholds for them. Russia has no gauges with thresholds: the Bank of Russia, hh.ru and HSE publish levels without a breakdown by cause.

The site's author proposes decisions that pay off on any path: measure the effect in accepted changes, not lines, and speed up review before generation (the longread on infrastructure for agentic development); hire juniors if the company is ready to pay for reproducing expertise (The Junior Engineer After Code); keep a retraining budget without expecting quick numbers: subsidised US programmes for the unemployed raise employment by 2–3 percentage points on average, and there is no data on corporate retraining of engineers.

What would change the picture is a combination: the BTOS share of firms and productivity above our thresholds at the same time, and a young-hiring signal that survives the controls. If that has not happened by early 2028, our slow-path hypothesis, not the authors' conclusion, still stands.