Retraining workers is the most popular policy answer to AI-driven job disruption, and Anthropic has now published an evidence review asking a plain question: do these programs actually work? The answer from 56 randomized US studies is "modestly, and not nearly enough if disruption is large." The report, "An Evidence Review of Subsidized Job Training," is coauthored by independent researcher David Roodman and Anthropic's Maxim Massenkoff, and appears as a working paper of The Anthropic Institute. A companion version is on arXiv, submitted September 7, 2026.
This is research from a top lab, but it is not about models. It is about what the evidence says on one of the main policy tools people reach for when AI automation comes up. This post explains the numbers, the one kind of program that beats the average, and what the authors suggest doing now. It sits alongside our earlier coverage of Anthropic's economic scenarios for AI and GDP and Anthropic's data on junior hiring and agents.
TL;DR: what the review found
| Question | Finding | Source |
|---|---|---|
| What was studied? | 56 randomized US job-training studies since 1973, plus European experimental evidence | Anthropic post, arXiv abstract |
| Average employment effect | Up two to three points per training slot offered (arXiv: 1.7 points in years 3 to 5) | Anthropic post; arXiv |
| Average earnings effect | About $1,000 a year (arXiv: roughly $800) | Anthropic post; arXiv |
| Cost | About $13,000 per person (arXiv: $13,598) | Anthropic post; arXiv |
| Does it pay for itself? | Roughly break even; government recovers more than half of its spend | Anthropic post |
| Biggest federal programs | Small or null effects (JTPA, Workforce Investment Act, Job Corps) | arXiv abstract |
| Best programs | Sector programs earn several times more, about ten times in the arXiv version, but are hard to replicate | Both |
| Conclusion for AI | Existing programs would likely fall short against large-scale displacement | Anthropic post |
What the numbers mean
The headline effect sizes are small. For each person offered a training slot, employment rises by two to three percentage points and earnings by roughly $1,000 a year, at a cost near $13,000. Anthropic notes that once added tax revenue and reduced benefit payments are counted, the government recovers more than half of what it spends, and the programs roughly break even overall. The arXiv abstract is more precise: it says they recoup about three-quarters of their cost for government and break even "under (debatable) assumptions about how long benefits persist."
Two details matter for reading these figures. First, the unit is "offered training," not "completed training," which is how randomized trials measure the policy people actually vote on, since not everyone offered a slot takes it. Second, the benefit stream is assumed to continue for years, so the break-even claim is sensitive to how long gains last. The authors flag this themselves.
Why the big federal programs disappoint
The arXiv abstract adds that evaluations of the largest federal efforts, namely JTPA, the Workforce Investment Act and Job Corps, return small or null effects. That matters for policy discussions that treat "retraining" as one thing. In the review's data, scale and generic design are exactly where average results come from, and they are not impressive.
The exception: sector programs
One cluster stands out. Sector programs partner with employers in a high-demand industry and place people directly into jobs in that field. According to the arXiv abstract, they screen applicants intensively, design curricula with employers and track local skill demand, and they raise earnings roughly ten times as much as the average. Anthropic's summary says gains are "several times larger."
The catch is replication. Attempts to copy the successful models "have often failed," Anthropic says. That tension is the core of the paper: the programs that work seem to depend on tight employer relationships, selective intake and local labor-market knowledge, none of which scale easily by copying a template.
Why this matters for AI
The review is part of Anthropic's Economic Research work on how AI affects the economy, which includes the Economic Index tracking AI use across occupations, a framework for measuring labor-market effects, and an Economic Policy Framework that lists possible responses including retraining. The authors' conclusion is direct: if AI displaces workers at scale, existing retraining programs would likely fall short.
That is a forecast built from historical programs, which were mostly designed for workers displaced by trade or plant closures, not for a technology shift that hits many occupations at once. It does not show that retraining cannot work for AI displacement, only that the best-evidenced version of it is small. Skeptics of AI's labor impact, such as the economist in our Acemoglu forecast post, would read the same evidence as a reason for caution about both the disruption and the remedies.
How to read a meta-analysis like this
A meta-analysis pools results from many studies to estimate an average effect. The appeal here is that every input is a randomized trial, the strongest design for separating a program's effect from who chooses to enroll. The limits are worth knowing before quoting the figures.
- Averages hide spread. An average of one to three points can combine programs that did nothing with a few that did a lot, which is exactly what the sector-program finding suggests.
- Old programs, new world. The trials span from 1973 onward. Labor markets, credentials and employer practices have changed, and none of the trials tested training for AI-era skills.
- Long-run persistence is assumed. The break-even result depends on how many years the earnings gain lasts, and the authors call those assumptions debatable.
- Offered versus completed. Results are per training slot offered, so they include people who dropped out. The effect for those who finish is larger than the headline.
- Publication bias. Meta-analyses can overstate effects if favorable studies are likelier to be published. We could not check how the authors handled this from the abstract alone.
What people are asking
Does this mean retraining is a waste of money? No. The review finds modest positive effects and roughly break-even economics on average. It says the typical program is not a strong enough tool for a large shock, and that a few designs are much better than typical.
What is a sector program? One built around a specific industry with employer input, with intensive screening of applicants, curricula designed with employers and tracking of local skill demand. Graduates are placed directly into jobs in that field.
Why not just copy the successful ones? Anthropic says replication attempts have often failed. The ingredients that seem to matter, such as selective intake and strong local employer ties, are hard to reproduce through a template.
Is this Anthropic's view of how AI will affect jobs? Not directly. It is an evidence review about training programs. It belongs to a broader research program on AI's economic effects, and it makes no claim about how many jobs AI will displace.
What the authors recommend
Their central recommendation is to invest now in demonstrating, evaluating and scaling the most promising programs. Concretely, that includes an effort to rapidly expand a leading program for a specific group of workers and to measure the results rigorously. Anthropic points to its Economic Futures Research Fund as designed to fund investigations into questions like these.
In practice this means three things for policymakers and funders:
- Back sector-style models rather than another generic voucher scheme.
- Build evaluation in from the start, using randomization where possible, because the replication failures are what the field most needs to understand.
- Pick a defined population for a fast scale-up, so success or failure is attributable.
What this does and does not tell builders and workers
If you build or buy AI systems, the useful reading is about timing. Policy support for displaced workers will not arrive from existing programs at the scale some forecasts imply, so internal reskilling inside companies and employer-linked training are likelier to matter in the near term. Our guide to AI courses and training programs is one way individuals can look at what skills are in demand, though the review's sector-program finding suggests the most valuable training is tied to a real employer pipeline, not a certificate alone.
For workers, the review is neither alarm nor comfort. Average training gains are positive and modest. The biggest wins come from programs that place people into specific jobs. That favors looking for training with employer partners and demonstrated placement outcomes.
Limits and open questions
We read Anthropic's post and the arXiv abstract, not the full text, so we cannot vouch for the study-by-study methods. Three caveats apply. The studies are US-centered, with European experimental evidence mentioned separately. Past programs were not designed around AI displacement. And the authors themselves call the persistence assumptions behind break-even debatable. Anthropic is also a party with an interest in how AI's economic effects are framed, so independent replication of the pooled estimates would be valuable.
The post date also deserves a note: the Anthropic page's metadata is dated August 12, 2026, and the arXiv version is dated September 7, 2026, while it surfaced in our feeds on October 22. We cite all three as reported on the pages.
Related reading
Figures are from Anthropic's research page and the arXiv abstract as of October 22, 2026.