Existing worker retraining programs would be unlikely to absorb a rapid rise in unemployment caused by AI, according to a new study from Anthropic. Independent researcher David Roodman and Anthropic economist Maxim Massenkoff find that training generally helps, but its average effect is too small to make a substantial difference during a large labor market shock.
The analysis combines 146 impact estimates from 56 randomized US studies conducted between 1973 and the present. Most tested programs aimed to move low-income adults and young people into occupations including IT support, nursing assistance and welding. They typically lasted around six months and cost approximately $13,000 for each person offered a place.
For programs in which training was the primary intervention, an offer of participation increased employment by 2.8 percentage points in the second year after randomization and by 1.7 points in years three to five. Annual pre-tax earnings rose by $1,139 in year two and $791 over the longer period, with all figures converted to 2025 dollars.
Those are statistically significant results, but not transformative ones. Writing on LinkedIn, Massenkoff described the effects as “small but positive impacts” and said it was difficult to see existing programs “meeting the moment during a surge of unemployment.”
Worker retraining has become one of the most widely supported proposed responses to AI-driven job losses. Anthropic’s Economic Policy Framework includes retraining among possible interventions, while its Economic Index and labor market research are intended to identify how AI use and exposure vary between occupations.
What 56 randomized trials actually found
The authors separate the full collection of interventions from 33 studies, covering 78 impact estimates, in which training was judged to be a primary component.
In the training-primary group, control participants had an average employment rate of 62.6% in year two. Offering the program increased that by 2.8 percentage points. In years three to five, the control employment rate was 63.4% and the estimated gain fell to 1.7 points.
The corresponding earnings increases were $1,139 in year two, from a control-group average of $14,645, and $791 in years three to five, from a base of $16,148.
The short-term results were somewhat stronger for employment, with a 4.8 percentage point increase during the first year. Short-term earnings were the one major result that did not reach statistical significance, potentially because time spent in training temporarily reduced the time available for paid work.
The findings are reported as intention-to-treat effects. In other words, they measure the impact of offering someone a place, whether or not that person completes the program. That is an important limitation, but also a realistic measure of what governments obtain when they fund access to training.
Among the training-primary experiments, 75% of people offered the tested program participated, compared with 9% of the control group. The offer therefore increased participation in that specific program by 66 percentage points.
The difference was much smaller when the authors counted participation in any training. Many control-group members found alternative courses, leaving an overall treatment-control gap of only 28 percentage points.
Depending on assumptions about the effectiveness of those alternatives, the impact per additional trainee could be two to four times the headline effect per person offered a place. The authors describe the upper end of that range as a likely upper bound. It does not alter the calculated benefit-cost ratios because both costs and effects rise when expressed per trainee.
Claude helped assemble the evidence, with methodological caveats
The paper is not presented as a purely systematic review. Roodman and Massenkoff say they used judgment to assess individual studies and program implementation before combining the US randomized evidence in a new meta-analysis.
Claude Opus was used, with human oversight, to extract program characteristics and impact estimates from thousands of pages of research. The model produced a reasoning table documenting its sources and logic, then conducted an adversarial review by returning to the original documents. The authors say they also spot-checked and revised the extracted data as the analysis developed.
The approach allowed the researchers to include considerably more US randomized evidence than previous meta-analyses, but it introduced its own judgment calls. Decisions included whether training was sufficiently central to an intervention, how follow-up periods should be classified and how results covering several months or years should be allocated.
Some government evaluations also reported only whether a finding passed a statistical significance threshold, rather than publishing a complete standard error. The researchers imputed p-values for those results, a decision they acknowledge influences the random-effects estimates.
The analysis is designed to represent the result of a typical study rather than the average effect across every person who participated in all 56 trials. The individual studies also tested different programs, populations and labor markets, limiting what comparisons between them can establish about cause and effect.
The economics depend on benefits lasting beyond the studies
The cost-benefit calculations are more encouraging than the employment figures, although they rely on assumptions extending well beyond most study periods.
For training-primary programs, the authors estimate an average upfront cost of $13,598 per treatment-group member. The projected net present value of additional pre-tax earnings is $14,146. When employer-paid taxes and other fringe benefits are included, the estimated increase in total compensation reaches $19,525.
That produces a narrow benefit-cost ratio of 1.44, or $1.44 in additional compensation for each dollar spent. A broader calculation, which accounts for lower spending on alternative training, raises the ratio to 1.80. The estimated societal internal rate of return is 5.97%.
Government does not recover its entire investment, but the calculations suggest that higher tax receipts, reduced public benefit payments and lower spending on other training could recoup around 76% of the original cost. The remaining long-term fiscal cost is estimated at approximately 24 cents for each dollar spent.
These figures should not be read as a straightforward cash return. Most evaluations followed participants for only a few years, while the economic model projects earnings effects through the age of 65. It assumes that gains decline slowly, drawing on separate research about the persistence of income changes.
The authors state that without this assumed persistence, the average training-primary program would have a benefit-cost ratio below one. The calculation also treats Social Security contributions as public revenue without counting the future benefits they may secure for workers.
Potential displacement is excluded as well. If a trained participant obtains a job that would otherwise have gone to another worker, the individual study can record a benefit without an equivalent increase in employment across the economy.
National US programs have not broken the pattern
Randomized evaluations of three large US workforce programs have produced results at or below the wider average.
The Job Training Partnership Act increased employment among low-income adults by approximately 2.3 percentage points in the late 1980s, from a base of about 70%. Annual earnings rose by around $1,100 in 2025 dollars, but researchers found no clear benefit for low-income young people.
Job Corps, an intensive residential program created for young people as part of the 1960s War on Poverty, produced no effect on employment or earnings in follow-ups extending to 20 years. The exception was a temporary employment increase of one to two percentage points in years three to five.
Anthropic reports that The Workforce Investment Act, which succeeded the Job Training Partnership Act, returned results that were approximately zero for both low-income adults and displaced workers. Estimates were as likely to be negative as positive and were not statistically significant.
That evaluation also shows how difficult it is to test a training offer when similar services remain available elsewhere. Many people offered the Workforce Investment Act training did not take it, while control participants found other programs. The difference in participation between the two groups was only 15 percentage points, leaving the study with limited power to detect an impact.
Sector programs produce the strongest earnings results
A smaller group of employer-linked “sector programs” performed considerably better. These programs concentrate on industries with current local demand, involve employers in designing curricula and combine occupational training with preparation for interviews and the workplace. They may also arrange internships, provide individual coaching, support participants after placement and maintain direct hiring pipelines.
Across the authors’ expanded collection of sector-program studies, the average annual earnings increase was $3,149 in year two and $3,703 in years three to five. Employment rose by around three percentage points in the medium term, but the long-term employment effect was only 1.3 points and was not statistically significant.
The advantage therefore appears most clearly in the quality or pay of work obtained, rather than in a large permanent increase in the number of people employed.
The average sector program cost $11,602 per person offered a place. Its projected lifetime impact on pre-tax earnings was $60,319, producing a narrow benefit-cost ratio of 7.18 and a broader ratio of 8.07.
The calculation suggests that an average sector program funded by government could generate more in additional tax revenue and reduced benefit use than it costs. The authors estimate a societal internal rate of return of 34.03%.
Individual results vary. Project QUEST, the Wisconsin Regional Training Partnership, Jewish Vocational Service in Boston and Per Scholas increased annual earnings by between $2,000 and $5,000 in early randomized evaluations.
Year Up produced an increase of around $8,000 a year, with no reduction in the effect after seven years. It was also substantially more expensive than the average program, at about $30,000 per participant.
Per Scholas was the strongest initial performer in the four-site WorkAdvance evaluation. Another participating organization, St. Nicks Alliance in New York, did not generate a clear early earnings gain, but its impact reached approximately $8,000 a year between seven and ten years after randomization.
The strongest programs admit a minority of applicants
Sector programs do not achieve those results by offering open access to everyone seeking work. The four organizations in the WorkAdvance evaluation admitted approximately one in five applicants after assessing qualities including motivation and commitment. In 2025, Per Scholas received 70,000 applications for 5,000 places. Its process included behavioral assessments, preparatory work and other tests intended to identify people capable of completing the program.
This selectivity helps explain why sector-program control groups were already doing better than participants in many other training studies. Their long-term employment rate was around 80%, compared with approximately 60% in the wider collection of training experiments. Depending on the follow-up period, sector-program control groups also earned between 50% and 75% more.
The programs are therefore not simply producing stronger results from an equivalent population. They are identifying a comparatively work-ready group that employers may otherwise overlook, then connecting those applicants with specific vacancies.
The report argues that this matching role is central. Sector programs monitor demand, adjust teaching around employer requirements and build enough trust for companies to consider candidates outside their normal recruitment channels. The approach replaces the “train and pray” model in which institutions deliver courses without a close connection to actual hiring demand.
A hiring commitment from employers was associated with earnings effects of approximately $8,000 to $9,000 in year two and beyond. However, that finding rests on only three programs: Year Up, the Wisconsin Regional Training Partnership and the earlier Wildcat program. It is promising evidence, not proof that adding a hiring promise will reproduce the same results elsewhere.
Replicating successful programs remains the unresolved problem
The evidence becomes much less reassuring when successful sector models are copied quickly. The Center for Employment Training in San Jose generated positive results in two early evaluations. It combined occupational teaching with a focus on better-paid work and relationships with employers, features later associated with the sector approach.
The US Department of Labor subsequently funded replications at 14 locations. Only four were judged to have reproduced the complete model with high fidelity, and none of the 14 sites produced a sustained, positive and statistically significant effect.
Evaluators pointed to CET San Jose’s unusually committed leadership, political advocacy and relationships built over two decades. The institution had developed within its local community. Reproducing its visible structure did not recreate those underlying relationships.
The same pattern appears elsewhere. Eight of the nine Pathways for Advancing Careers and Education evaluations produced no clear employment or earnings effects. Year Up was the exception.
In WorkAdvance, the established Per Scholas program outperformed the three newer sites during the initial follow-up. St. Nicks Alliance eventually recorded a large earnings gain, but only after several years.
The research supports a distinction between deciding to create a sector program and successfully operating one. Employer collaboration, market knowledge and institutional trust take time to build. A rapid national rollout could copy the program format while losing the capabilities that made the original effective.
UK and European evidence is considerably thinner
High-quality evidence outside the US is limited, and several European experiments encountered problems that make their findings difficult to interpret.
A Norwegian trial covering 770 applicants reported earnings approximately 17% higher in the two years after training. However, the treatment and control groups differed substantially before the program began, suggesting that the randomization may not have worked as intended.
A Danish experiment involving 810 people found possible negative effects following an average two-week course. Almost half of the treatment group did not participate, while 22% of the control group accessed the training. The small difference in take-up made modest effects difficult to detect.
In the Netherlands, 160 people receiving unemployment benefits were divided between a control group, training vouchers and a short program intended to build confidence in job searching. Employment reached 26% in the latter group after six months, compared with 9% for vouchers and 11% for the control group. Attrition, crossovers and the exclusion of some participants left those figures exposed to substantial selection bias.
The UK Employment Retention and Advancement demonstration provides a larger randomized evaluation, but training was only one component of a package that also included employment advice and financial incentives.
Among people aged 25 and over who had experienced long-term unemployment, the overall package increased employment by approximately two percentage points and earnings by £296 a year over five years. Only around 10% used the training incentives, preventing the evaluation from showing whether training produced the result. The authors conclude that the UK study provides little direct evidence that government training offers helped.
The evidence does not yet match an AI displacement scenario
The central limitation is straightforward: none of these studies tests retraining following job displacement by large language models. Most participants were low-income adults, young people or those with a long history of unemployment. Only around 3% to 4% of the studies focused on displaced workers. Nearly all the programs lasted less than a year.
The analysis also excludes apprenticeships, community colleges, four-year degrees, professional education and online learning. It therefore covers only one part of the system through which workers acquire new skills.
That evidence may still apply to people displaced from lower-paid jobs who can move into an occupation requiring several months of training. It is a poorer fit for accountants, lawyers, software developers and other professionals who may need several years of education to approach their previous earnings.
Longer training creates a further forecasting problem. Sector organizations can respond to current demand for an entry-level technician and prepare someone within six to 12 months. They will have much less certainty about which occupations employers will need four years later, particularly if AI capabilities and hiring requirements are changing quickly.
The paper identifies Trade Adjustment Assistance as one possible model for people who can demonstrate that they lost work because of AI. The US program combined funding for training with extended unemployment insurance, job-search assistance and relocation support. A quasi-experimental study estimated an additional 20 months of work and $50,000 in earnings over ten years, but it could not separate the effects of training from those of income support.
A trigger-based scheme would also miss workers affected through reduced hiring rather than direct layoffs. If companies replace departing employees with AI systems, or expand without recruiting into exposed roles, no individual dismissal necessarily activates the support.
Roodman and Massenkoff’s recommendation is to invest before an employment emergency. They propose a “fire drill” in which a leading sector program is rapidly expanded for a defined group of workers and evaluated through randomization. That would test whether its employer relationships, selection process and outcomes survive growth.
Anthropic says its Economic Futures Research Fund is designed to support investigations of this kind, with the immediate research priority focused on demonstrating, evaluating and scaling the most promising workforce programs.
Leave a Reply