Retraining displaced workers is, in the words of Anthropic's announcement, "the most popular policy option for mitigating labor market disruption from AI." This review asks whether that option would work if AI did displace workers at scale, by looking at how well subsidized job training has worked in the past. It was published on 12 August 2026 as a short post on anthropic.com (headed "Reviewing the evidence on worker retraining programs" and indexed as "How well do job retraining programs work?") with a 124-page report, An evidence review of worker retraining, numbered The Anthropic Institute Working Paper 2026-01. The authors are David Roodman, an independent researcher, and Maxim Massenkoff of Anthropic. The post's BibTeX entry gives a fourth title, "An Evidence Review of Subsidized Job Training."
The report is the version of record for every number in this note; the post is a short summary that rounds some figures differently (see "The post and the report"). The report's code and data are in a public repository (github.com/droodman/job-training-meta-analysis).
What the review does
Scope and evidence standard
The review is mostly about training programs for job seekers in the United States, typically low-income adults and young people trained over months for jobs such as nursing aide, IT support technician or welder (Executive summary). Of the kinds of workforce help the report lists, it covers soft-skills teaching, adult basic education, vocational teaching and on-the-job training; job-search help, subsidized employment, apprenticeships and public employment are mostly left out (§2.1). The authors note that apprenticeships are large (678,014 active US apprentices in 2025) but that they found little high-quality evidence on them.
The authors describe their approach as "not purely a systematic review" (Overview). It combines:
- A high evidentiary standard. Randomized trials are favored, plus a few natural experiments with a strong claim to arbitrary assignment (regression discontinuity and "judge randomization", in which caseworkers or investigators with different tendencies are assigned as good as at random). Matching studies are excluded, though footnote 3 concedes that some may be as credible as studies included.
- An aggressive search, including reports not previously online, found at the Library of Congress and the Department of Labor library (footnote 4). It yielded 56 randomized US studies, 7 randomized or discontinuity-based studies in Europe and two judge-randomization studies, one American and one Danish (Overview).
- A new meta-analysis of the 56 US studies (§6.5).
- Case histories of the four major federal evaluations (§4), the quasi-experimental studies (§5) and the "sector programs" (§7).
Two measurement choices recur throughout. Most estimates are intention-to-treat (ITT) effects: the difference in outcomes between people offered a program and people not offered it, whether or not they took part. The local average treatment effect (LATE) divides the ITT by the difference in participation and estimates the effect on people whose participation the offer changed (§3.2). The report defaults to the ITT because it is easier to define and, the authors argue, no less likely to fit a future disrupted labor market.
The meta-analysis and Claude's role in it
The meta-analysis pools 146 impact estimates from the 56 US studies, randomized between 1973 and the present; 24 studies report results for several groups or sites (§6.5.1). Studies were drawn from six earlier reviews and recent follow-ups. Because most evaluations are long reports to government agencies, the authors used Claude Opus (version not stated) "with engaged oversight" (§6.5) to extract study traits and impact estimates from "thousands of pages of reports." Claude was instructed to produce a "reasoning table" of sources and logic for each value and then to run an adversarial review of its own extractions against the sources; the authors spot-checked and revised the extraction as they analyzed the data (§6.5). Claude also judged which interventions were "training-primary", meaning training was a primary feature: 33 of the 56 studies, with 78 estimates, cleared that bar (§6.5.1).
The report gives no error rate for the extraction and no count of spot checks. It quotes an earlier meta-analysis's warning that extraction involves "measurement errors and errors of interpretation", and says it applies "as well to this AI-accelerated effort" (§6.5). Where studies report only significance thresholds, p-values are imputed (0.005, 0.03, 0.075 and 0.55 for the four usual bands), and the authors say these "uncertain imputations influence the meta-analytic, random-effects results" (§6.5).
Impacts are pooled with random-effects meta-analysis (restricted maximum likelihood), with standard errors clustered by study, so the averages describe the typical study rather than the average participant (§6.5.2). Follow-up is split into year 1, year 2 ("medium-term") and years 3 to 5 ("long-term"). All dollar figures are in 2025 dollars.
What it finds
Average effects are positive and small
From Table 2 (§6.5.2), the pooled ITT effects are:
| Outcome | All 56 studies | Training-primary studies |
|---|---|---|
| Employment, year 2 | +2.6 points (SE 0.4) | +2.8 points (SE 0.6) |
| Employment, years 3 to 5 | +1.8 points (SE 0.5) | +1.7 points (SE 0.5) |
| Earnings, year 2 | +$984 (SE 255) | +$1,139 (SE 446) |
| Earnings, years 3 to 5 | +$657 (SE 185) | +$791 (SE 302) |
Control-group employment rates are about 56% to 63% depending on sample and period, so the year-2 effect is about 2.6 to 2.8 points on a base of about 60% (§6.5.2). The report describes the earnings effect as a 7.8% rise in year 2 and 4.6% in years 3 to 5. The report says eleven of the twelve employment and earnings estimates in Table 2 are "2–4 times their standard errors" (by this note's arithmetic all eleven are at least twice their standard errors, and the year-2 employment estimates are 6.5 and 4.7 times theirs); the exception is the year-1 earnings effect of training-primary programs (+$835, SE 553), which the authors say "may reflect temporary earnings losses while in training." The executive summary leads with the longer-run figures: 1.7 points of employment against a control base of 63%, and $800 a year.
Low participation dilutes these figures. In training-primary trials, people offered a place were 66 percentage points more likely to take the targeted program but only 28 points more likely to receive any training, because many control-group members found training elsewhere (§1.4, §6.5.2). Scaling the ITTs by those differences gives effects about 1.5 times larger per program participant and nearly four times larger per person trained; the report calls the per-trainee figure "likely an upper bound" (§1.4) and notes that it relies on the alternative trainings being equally effective (§6.5.2).
Benefits and costs
Among the 67% of training-primary interventions that report costs, the average cost is $13,598 per person offered training (§1.2). Projecting earnings gains to retirement, which requires assuming how long gains last, gives a net present value of $14,146, or $19,525 with employer-paid taxes and fringe benefits (§1.2, Appendix G). The authors say this "should be read as a little better than break-even" (§1.2). Their benefit-cost summary for training-primary programs (§6.5.3) gives a narrow benefit-cost ratio of 1.44, a broader one of 1.80, a marginal value of public funds of 2.52, and a societal internal rate of return of 5.97%. If government pays, programs recover about 76% of their cost through higher tax revenue and lower benefit use, so the long-term fiscal cost is about 24% of the upfront cost; footnote 6 says the payback is "substantially lower" if Social Security contributions are treated as buying future benefits rather than as taxes.
Large federal programs
The US government commissioned randomized evaluations of three national programs (§1.3, §4):
- Job Training Partnership Act (JTPA), late 1980s: for low-income adults, employment rose about 2.3 points from a base near 70% and pay about $1,100 a year; there was no clear benefit for youth.
- Job Corps, a residential program for young people: no effect on employment or earnings over follow-ups of 20 years, apart from a temporary 1 to 2 point employment rise in years 3 to 5.
- Workforce Investment Act (WIA), JTPA's successor: estimates near zero for low-income adults and dislocated workers. Uptake was a problem; the treatment-control difference in participation was only 15 points.
The report's discussion of the JTPA (§3.5) describes performance standards that led some local agencies to delay recording enrollments and that gave them "a powerful incentive to engage in 'creaming'", serving applicants likely to find jobs anyway; it uses the case as an example of how incentives can distort a program at scale.
Dislocated workers
Evidence on workers displaced by broad economic change is "sparser and mixed" (§1.5). A Danish study that exploits as-good-as-random caseworker assignment finds that an average of 52 days of classroom training helped people change occupations, while on-the-job training showed no such effect (Humlum, Munch and Rasmussen, 2025, as reviewed in §5.2). A US study of Trade Adjustment Assistance, with what the authors call a weaker claim to arbitrary assignment, finds that eligible workers later worked and earned more, but training cannot be separated from the extended unemployment insurance the program also provided (Hyman, 2018, §5.1). Randomized pilots in New Jersey and Texas found training plus job-search help no better than job-search help alone at reducing unemployment within 12 months (§1.5). Programs backed by an employer's commitment to hire raised earnings "an order of magnitude" more than average, a result driven by three programs (§1.5, §6.5.5).
Sector programs
"Sector programs" screen applicants, work with employers on what to teach and place graduates in specific high-demand industries; examples include Year Up, Per Scholas and Project QUEST (§1.6, §7). As a group in the meta-analysis they cost about the same as other training-primary programs ($11,602 against $13,598 per person offered), but the present value of their earnings effects is $60,319 against $14,146; the report gives their benefit-cost ratio as 7x (§1.6; 7.18 in §7). The ratio of the two pre-tax figures is about 5.2 by this note's arithmetic, so the 7x presumably counts total compensation, as the narrow benefit-cost ratio of §6.5.3 does (for training-primary programs, $19,525 / $13,598 gives that ratio's 1.44). The executive summary says some raised pay by $5,000 to $10,000 a year, about ten times the average. The report credits several features, while saying more research is needed to tell which matter (§1.6, §7.4):
- Selectivity. One evaluation of four programs reports admitting about one in five applicants (Hendra et al., 2016, as cited); Per Scholas told the authors it had 70,000 applicants for 5,000 places in 2025 (§7.4, footnote 56, a conversation with staff).
- Short courses that teach skills learnable in months.
- Close ties to employers, who in Year Up's case fund much of the training and commit to hiring.
- Private operation, which makes screening and curriculum changes easier.
Replication is the main caveat. A Department of Labor replication of the Center for Employment Training (CET) in San Jose failed at all 14 sites, including four run by CET itself outside San Jose (Miller et al., 2005, as cited in §1.6 and §7.1.1). Only two replications of the sector approach have raised incomes in randomized trials: Year Up's expansion beyond Boston and one of three new WorkAdvance sites, St. Nicks Alliance in New York (§1.6). Of the three promising programs deliberately scaled or replicated under randomized evaluation, the authors count only Year Up as showing consistent gains (§3.5).
Lessons for an AI-disrupted labor market
Section 8 draws four conclusions, which the authors present as extrapolations "into a highly uncertain world" (Executive summary):
- Short programs of the kind reviewed could help some workers in lower-wage jobs displaced by AI (§8.1).
- They are "ill-matched" to more skilled workers such as accountants, lawyers and programmers, whose reskilling may take years and whose future demand is hard to forecast (§8.2). The report cites a Danish study in which injured skilled-trade workers who went to college eventually earned 25% more, after a typical four years of study.
- A Trade Adjustment Assistance style approach, which buys time by extending unemployment insurance, is "worth considering" for people demonstrably fired because of AI, though it would not help when firms simply stop hiring into a role (§8.3).
- Public and private funders should "aggressively support demonstrations, evaluations, and scaling of sector programs", for example through a "fire drill" that rapidly scales one program for a targeted group and evaluates it (§8.4).
The overall judgment is that current programs "may be a defensible government expenditure" but have effects "small enough that they would not leave a dent in a persistently high unemployment rate" (§9).
The post and the report
The announcement post and the report agree in substance but round differently:
- The post's "two to three percentage points" of employment and "roughly $1,000 a year" match the report's year-2 (medium-term) estimates; the executive summary's headline figures (1.7 points and $800) are the years 3 to 5 estimates.
- The post says the government "recovers more than half of what it spends"; the report's estimate is about 76% (§6.5.3).
- The post describes sector programs' gains as "several times larger"; the executive summary says some raise earnings "10 times as much", and §1.6 gives a 7x benefit-cost ratio for the group.
The PDF's text layer (not the rendered page) carries the line "Anthropic Economic Index report: Cadences" at the foot of page 1, apparently left over from another document's template; it shows up in text extracts and has no bearing on the content.
Limitations and open questions
The authors state most of these themselves:
- External validity. Nearly all the evidence concerns short programs for low-income or long-unemployed people. A wave of displaced white-collar workers would be, in the Overview's words, "the most educated and highly paid in history", with more savings and different needs. Only 3 to 4% of the interventions in the meta-analysis are meant for dislocated workers (§6.5.1, text after Table 1).
- Correlates are not causes. Comparisons across studies (which program features go with larger effects) are "not as dispositive" as the individual trials (§6.5.4).
- Projected benefits. The benefit-cost results depend on assumptions about how long earnings gains last beyond the few years most studies observe (§1.2). Footnote 44 adds that successful programs may be followed longer than others.
- Attrition. About half of the sector program evaluations rely on surveys rather than administrative wage records, which leaves them open to attrition bias (§3.3).
- AI-assisted extraction. Study traits, impacts and the training-primary label were extracted by Claude Opus with human spot checks; no extraction accuracy is reported (§6.5). Imputed p-values also affect the pooled results.
- Scaling and selection. The best results come from selective, privately run programs whose organizational ties took years to build; scale-ups and government-run programs tend to do worse in the literature the report cites (§3.5).
- A living repository. The code and data repository has been updated since publication. A commit of 29 October 2026 (d826c39) is titled "Fix: impute 0.005 instead of 0.01 for ***", the value §6.5 states, which suggests (this note's inference) that the code behind the published numbers used 0.01. The same commit changes the repository's benefit-cost input file: the training-primary earnings effect becomes $1,156 in year 2 (published $1,139) and $798 in years 3 to 5 (published $791), and the sector-program year-2 effect $3,192 (from $3,149). This note follows the published report; it did not rerun the analysis or check other tables.
The report has no conflict-of-interest statement. Anthropic, Massenkoff's employer, states in the post that its Economic Futures Research Fund "is designed to fund investigations into questions like these."
Research questions
- How accurate was Claude's extraction? Re-coding a random sample of the 146 estimates by hand from the original reports and comparing values and training-primary labels would give an error rate the report does not state, and show whether errors cluster in particular outcomes or timeframes.
- How far do the October 2026 repository changes move the published results? The p-value imputation fix shifts the training-primary earnings inputs by about 1% to 2%; rerunning the full analysis at the release commit and at the latest commit would show whether any Table 2 figure, benefit-cost ratio or meta-regression changes materially.
- Would sector-program effects survive a rapid scale-up? The proposed "fire drill" is the test: a randomized evaluation of one program expanded quickly for a defined population, compared with the original site's effects, would show how much of the gain depends on slow-built employer ties.
- Do training effects differ for displaced professionals? Randomized or quasi-random evaluations of retraining for college-educated workers, such as compressed degree programs, would show whether the low-income evidence base transfers at all.
- How much of a sector program's gain comes from selection rather than training? Studies that randomize among already-screened applicants measure the training effect on selected people; varying the screening while holding the training fixed would separate the two.
- What share of AI-related job loss would trigger-based aid reach? The report notes that Trade Adjustment Assistance style aid does not fire when firms stop hiring rather than lay off. Data on whether AI exposure shows up in layoffs or in reduced hiring would show how much of the problem such aid could address.
How it connects
- The post places the review alongside Anthropic's framework for measuring labor market effects, Labor market impacts of AI: A new measure and early evidence, which the report also cites (Massenkoff and McCrory, 2026) for the view that, on current usage data, college-educated workers are most at risk. The report's opening statement that job loss is Americans' top concern about AI cites Anthropic's Results from first Anthropic Public Record.
- The post links Anthropic's Economic Policy Framework (June 2026), which it says sets out possible responses "including worker retraining", and the Economic Futures Research Fund. An earlier Anthropic policy post, Preparing for AI's economic impact: exploring policy responses (October 2025), listed workforce training grants among ideas for study; this note makes that connection, which the review does not cite.
- Scenarios for our Economic Future, published a month later by Anthropic's economics team, models how many cognitive workers would need to change occupations under different AI scenarios and names retraining among the mechanisms that would then matter; this connection is the note's own.
- Can we predict the jobs robots will do? (September 2026), with Massenkoff as second author, measures which jobs robots can already do; it does not cite this review, and the connection is the note's own: the review's §8.1 discusses drivers as workers short programs might help.
References
- Roodman, D., and Massenkoff, M. (2026). An evidence review of worker retraining. The Anthropic Institute, Working Paper 2026-01. Post (12 August 2026): https://www.anthropic.com/research/reviewing-the-evidence-on-worker-retraining-programs. Report PDF (124 pages): https://assets.claude.com/169e4adceebbab066b91e0b45127d73cb2bda24b.pdf. Section numbers in this note follow the PDF. The same report was posted to arXiv as 2609.07011 (v1, 7 September 2026, 125 pages); a word-level comparison found only layout differences, with the same numbers. record
- Code and data: https://github.com/droodman/job-training-meta-analysis (last commit before publication: ccb8a6b, 9 August 2026).
- Studies cited above only as the report describes them (among them Hendra et al., 2016; Humlum, Munch and Rasmussen, 2025; Hyman, 2018; Miller et al., 2005; Karger et al., 2026) are listed in the report's references and were not read for this note.