What the New York Fed Retraining Paper Shows About Career Moves

What should a worker conclude from the paper?

The practical conclusion is narrow: training can be one route after work in an AI-exposed occupation, but the report does not support treating an AI-intensive pivot as the default. The Fed in Print record summarizes positive short-run returns for trainees, while the report finds lower returns among participants targeting more AI-intensive occupations than peers pursuing general training. These findings can coexist: a group can gain against its comparison group and still gain less than another training group. Keep three quantities separate. A positive return against matched controls is not a promise of a raise. A lower return for AI-intensive targets does not show every AI course fails or broader training always wins. The report’s estimate that 25%–40% of occupations meet its AI Retrainability Index definition describes an occupation-level classification, not the share of workers who will succeed or anyone’s probability of losing a job. For a current worker, compare a task-level upgrade or adjacent role with a larger change, using local demand and real constraints. The paper is a reason to ask what training is for, not to discard experience or buy a credential because an occupation is exposed.

Sources: How Retrainable Are AI-Exposed Workers? | Fed in Print; How Retrainable Are AI-Exposed Workers?

What population and comparison does the paper actually study?

Staff Report 1165 studies U.S. participants in Workforce Investment Act and successor Workforce Innovation and Opportunity Act programs who exited between 2012 and 2023. The dataset covers more than 1.6 million occupational-training participation spells linked to earnings and occupation records. The authors use occupations before training, those targeted by training, and post-training occupations to classify transitions against occupational AI-exposure measures. Because the observation window includes earlier forms of automation as well as later language-model exposure measures, the findings describe historical program experiences, not a direct trial of a current AI course. (“How Retrainable Are AI-Exposed Workers?”) The main comparison is between Title I participants who received occupational training and matched Title III participants who received job-search assistance without enrolling in that training. Administrative records provide participant characteristics, service history, quarterly earnings, and occupation codes. The researchers match trainees with job-search-only participants who have similar measured characteristics, including similar occupational exposure in the same year. Both groups receive job-search support, so the comparison estimates the additional association of occupational training. Matching also addresses the possibility that a pre-training earnings dip could make later recovery appear to be a training gain. The authors compare pre-training earnings patterns and use nearest-neighbor matches to reduce this concern. (“How Retrainable Are AI-Exposed Workers?”) This design is observational, not randomized. Caseworkers help determine which participants enter Title I training, and the report acknowledges possible selection into training rather than job-search-only services. Matching can balance recorded characteristics but cannot remove every unmeasured difference in motivation, circumstances, or caseworker judgment. The estimates are therefore group comparisons, not experimental proof that training caused each earnings difference. Occupation codes also serve distinct analytical roles: pre-training occupation identifies whether someone came from more-exposed work; the targeted occupation indicates whether the intended destination was more AI-intensive; and post-training occupation records where the worker was later observed. This allows separate questions about exposure at origin and intensity at destination. It does not isolate the effect of a specific course or credential. (“How Retrainable Are AI-Exposed Workers?”) The sample is unlike an employed professional deciding whether to study after work. WIA/WIOA serves people seeking workforce services. Title I includes adult and dislocated-worker programs, the latter generally serving people laid off or facing termination. Complete earnings and occupation data narrow the final matched analysis to 109,038 training spells, of which 108,215 were successfully matched to controls. The earnings analysis focuses on people with positive earnings in each of six observed quarters: three before and three after participation. This helps compare earnings among reemployed participants, but does not measure every participant’s chance of finding work; many records also lack complete occupation codes. An incumbent analyst weighing a self-funded evening course faces different selection, costs, time limits, and employer conditions. The report can inform questions about workforce training and destination occupations, but cannot forecast that person’s return. (“How Retrainable Are AI-Exposed Workers?”; “How Retrainable Are AI-Exposed Workers? | Fed in Print”)

Sources: How Retrainable Are AI-Exposed Workers?; How Retrainable Are AI-Exposed Workers? | Fed in Print

What does the earnings result say—and what does it not say?

For the participants who entered training from occupations above the study’s median AI-exposure score, How Retrainable Are AI-Exposed Workers? estimates average quarterly earnings returns of about $1,470 relative to matched participants who received job-search assistance without training. The report says this estimate is about 25% below the return for trainees from lower-exposure occupations. Those are two different comparisons: $1,470 is the estimated difference from a control group for the high-exposure-origin trainees; 25% describes how that group’s estimated return compares with the low-exposure-origin group. Neither number is an individual forecast, a guaranteed raise, or the return on a particular course.

The study’s outcome window is short. Researchers linked workforce-program records to state earnings records for three quarters before program entry and three quarters after exit, and report quarterly earnings in 2010 dollars. The analyzed sample is narrower than the headline database: it includes people with positive earnings in all six observed quarters, recorded occupations throughout, and the matching covariates. Requiring positive earnings focuses the analysis on reemployed participants and separates earnings changes from the chance of becoming reemployed at all. It therefore does not tell a reader the full effect of training on employment, nor does it capture long-run earnings progression, later job stability, or the return to an evening course taken while already employed.

The comparison is also observational. The authors match trainees to similar program participants who received job-search assistance but did not enroll in occupational training, using recorded characteristics and earnings patterns. Matching helps address measured differences, including the pre-training earnings dip discussed in the report, but participants were not randomly assigned to training. Caseworker decisions and unobserved differences may still affect who trains, which path they choose, and their later earnings. The estimate is evidence about a defined group under this design, not proof that training alone caused the entire gap.

Time matters to the result. The report says positive estimated returns across groups were largely driven by the most recent years in its 2012–2023 program data, when labor markets were tightest. Its interpretation is that training may carry stronger signal value when firms reach further into the skill market. That is the authors’ explanation of a historical pattern, not evidence that a future tight market will produce the same return or that training has this effect in every cycle. Hiring conditions may shape whether a participant finds work and how employers read training, so the estimate should not be detached from its period.

For a worker deciding whether to pay for retraining, the practical lesson is narrower than “training pays.” First identify a specific task or qualification gap in a role you could realistically pursue. Then look for repeated requirements in relevant local postings and check prerequisites, training time, total cost, salary needs, and whether the learning includes feedback or supervised practice. The New York Fed paper does not evaluate a named credential or establish that a course will close a particular gap. Its $1,470 figure cannot be converted into a personal salary expectation or tuition budget. Use it as a reason to investigate plausible paths, while requiring evidence about your target work before making an expensive commitment.

Sources: How Retrainable Are AI-Exposed Workers?; How Retrainable Are AI-Exposed Workers? | Fed in Print

Does the paper favor AI specialization or broader training?

Among participants who entered workforce training after leaving high-AI-exposure occupations, the New York Fed paper reports positive estimated earnings returns for both people targeting high-AI-intensity occupations and peers pursuing more general training. The AI-intensive-target group’s estimate was about $1,040 per quarter, roughly 29 percent below the comparison group’s return. That is evidence against treating AI specialization as the default next move. It is not evidence that AI training produces losses: the estimate remained positive relative to matched participants who received job-search assistance without occupational training. Those figures answer a narrow comparison. The researchers followed U.S. WIA/WIOA program participants from 2012 through 2023, linked training records to occupational codes and quarterly earnings, and matched trainees with similar job-search-assistance recipients. Within the high-exposure-origin group, they distinguished trainees by whether their target occupation was high or lower in the paper’s AI-intensity measure. In “How Retrainable Are AI-Exposed Workers?,” the authors describe the comparison as returns to AI-deepening versus more general skill investments. They do not compare a named machine-learning degree with a particular certificate, vendor course, or workplace project. Nor were trainees randomly assigned to target occupations. People’s backgrounds, chosen programs, caseworker decisions, local opportunities, and hiring conditions can differ in ways a matching design does not fully remove. The reported gap therefore should not be read as the causal effect of buying an AI course. Target occupation matters because a training outcome combines what someone learns with where they try to use it. A course may teach relevant techniques, while the destination role has its own entry requirements, competition, location limits, and demand. The report’s outcome is quarterly earnings for a defined participant group over a short post-training window; it does not isolate whether any particular skill was useful, whether employers adopted AI broadly, or how an incumbent employee’s evening study would fare. The authors also report that positive returns across groups were largely driven by the latest years in their sample, when labor markets were tightest. That historical pattern makes the hiring context part of the interpretation, not a forecast that the same return will hold in a different market. Imagine a claims analyst whose work includes drafting routine summaries, checking exceptions against policy, explaining decisions, and remaining accountable for a correct file. One possible move is to learn an approved tool for a bounded drafting task, then verify outputs against policy and track where corrections remain necessary. Another is to investigate a nearby role that places more weight on exception handling, policy interpretation, or quality review, building on knowledge already used in the current job. A larger AI-specialist transition is also conceivable, but it should be compared with actual target-role requirements and the analyst’s preparation, time, cost, location, and income needs. This is an illustration of how to frame alternatives, not an outcome observed by the study. The practical reading is conditional: keep AI-specific training on the list when a defined target role repeatedly appears in the reader’s reachable labor market and training closes a verified prerequisite gap. Otherwise, test whether a workflow upgrade or adjacent move makes better use of existing expertise before committing to a larger transition. The paper supports comparing destinations; it does not rank a reader’s personal options or establish that broader training is always superior.

Sources: How Retrainable Are AI-Exposed Workers?

What does the AI Retrainability Index measure?

The 25%–40% figure is an estimate about occupations in the New York Fed report’s historical training sample. The “AI Retrainability Index” ranks occupations by the share of their workers who, after leaving, moved into occupations with both higher pay and greater measured AI intensity. It does not say that 25%–40% of exposed workers will retrain successfully, and it is not an individual’s chance of keeping or losing a job. In “How Retrainable Are AI-Exposed Workers?”, the estimate comes from linked U.S. WIA/WIOA program and earnings records covering 2012–2023, with trainees compared to matched participants receiving job-search assistance without training. The two conditions in the index matter. A move toward an occupation with higher AI intensity alone could reflect a broad change in which occupations are hiring, rather than a worker being rewarded for new skills. The report therefore also asks whether workers’ earnings rise. Its authors explain that defining successful retraining through both skill movement and earnings helps mitigate survivorship bias: the index is less likely to label a transition successful just because AI-intensive firms have become more prominent in hiring. The higher-pay condition makes the measure about transitions that carried an earnings gain in the observed data, not simply movement toward a higher exposure score. The report further separates earnings gains while measured AI skills stay constant from gains associated with transitions toward more AI-intensive work. The first component may include wage increases for reasons such as occupational licensing barriers; the second reflects a change in the measured skill space. This decomposition is useful because an earnings improvement after a training spell does not, by itself, show that AI instruction caused the improvement. The index describes patterns of occupation changes and pay in the matched sample; it does not isolate the effect of a particular course, credential, or curriculum. “AI-intensive” also refers to an occupation-level measure of potential task relevance, not verified AI use in every workplace. The report merges occupational exposure measures with workers’ recorded occupations. It notes that such measures capture potential rather than realized substitution. A high index classification therefore cannot tell you whether a specific employer has adopted a tool, redesigned a role, or expects a worker to use AI. Nor does the index establish that an occupation is safe, doomed, or easy to enter. For a career decision, read the index as a historical clue about transitions from an occupation among program participants whose records could be analyzed. It can prompt a useful question: what nearby, higher-paying roles might use more AI-related skills? It cannot answer whether you personally should pursue one. Check the role’s actual tasks, prerequisites, local openings, likely pay, and training costs against your constraints. Treat the index as a reason to investigate a path, not a ranking of individual prospects.

Sources: How Retrainable Are AI-Exposed Workers?

How should exposure be separated from adoption and displacement?

Exposure measures whether work tasks may be relevant to AI capabilities; it does not show that an employer has adopted AI, changed a role, reduced hiring, or displaced a worker. Adoption, redesign, labor demand, and displacement require different observations. The New York Fed retraining paper groups workers partly by occupational exposure. That grouping alone cannot show that their employers used AI, their tasks changed, or AI caused a separation. The International Labour Organization’s “Generative AI and Jobs: A Refined Global Index of Occupational Exposure” shows why exposure is task-based. Published in May 2025, its global index combines task data, worker perceptions, expert input, and model predictions. The research drew on a representative sample of 29,753 tasks in Poland’s occupational classification, a survey of 1,640 employed people across broad occupational groups, and 52,558 assessments covering potential automation for 2,861 tasks. Its framework assigns four exposure gradients. The ILO estimates that one in four workers globally is in an occupation with some exposure and 3.3 percent of global employment is in the highest gradient. It concludes that transformation is more likely than full automation because most occupations contain tasks requiring human input. These are modeled exposure assessments, not counts of jobs automated or forecasts of people losing work. A separate New York Fed source illustrates adoption evidence. “AI and the Labor Market: Will Firms Hire, Fire, or Retrain?” reports August 2024 regional business surveys from New York and northern New Jersey. Among surveyed service firms, 25 percent said they had used AI in the prior six months; among manufacturers, 16 percent did. These are reported firm-use shares in two regional surveys, not national adoption rates for exposed occupations. Among firms already using AI, 10 percent of service businesses reported reducing worker counts in response and 5 percent increasing them; manufacturers reported no count changes. About a quarter of service AI users and 31 percent of manufacturer AI users said they had trained workers to use it. These are self-reports from a regional sample, not a causal estimate. Plans need their own label. Of firms planning future AI use, 53 percent of service businesses and 47 percent of manufacturers expected to retrain workers in the next six months. The authors also report planned hiring and limited planned layoffs, while cautioning that expectations could reflect overoptimism or tight labor markets. An intention is not a completed training, hire, or layoff. A survey snapshot does not establish what later happened. Together, the sources show why exposure labels cannot settle what is happening at work: task potential, firm use, planned change, and realized staffing outcomes are distinct evidence. For a career decision, inspect recurring tasks in your role. Which are being trialed or changed? Who checks the output, and who remains accountable when it is wrong? A report of use does not establish that a whole role is automated; continued human review does not prove every task is unchanged. Look for concrete workflow changes and confirmed staffing decisions, while treating forecasts as plans. This local picture is more useful for choosing a proportionate next step than inferring adoption or displacement from an occupation’s exposure score alone.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure; AI and the Labor Market: Will Firms Hire, Fire, or Retrain?

What do newer labor-demand signals add to the paper?

Newer evidence adds context about labor demand and employer behavior, but it does not revise the New York Fed’s 2025 training estimates. The studies observe different things: one tracks advertised vacancies by occupational AI exposure, another summarizes regional firms’ reported use and plans, and a third evaluates WIOA participation with a different automation measure. None establishes what will happen to a particular worker. In “Do Job Postings Show Early Labor-Market Effects of AI?”, published in May 2026, the New York Fed combines Lightcast U.S. job listings with an occupation-level exposure measure built from O*NET tasks and observed AI use. It weighs capability, observed usage, and automation versus augmentation, then aggregates task scores by occupation. This is grounded in observed use, but remains a measure of potential exposure, not proof that employers adopted the technology or cut jobs. The authors use an event study to compare vacancy trends in occupations above and below an exposure threshold around ChatGPT’s late-2022 release, accounting for occupation and economy-wide changes. They report that postings for more-exposed occupations had already declined relative to less-exposed occupations before the release, with no clear additional break afterward and a gap that stabilized after 2023. Within highly exposed occupations, the study also finds no clear divergence between junior and senior postings. These results complicate a simple story in which ChatGPT caused the relative decline or uniquely reduced advertised entry-level demand. Postings are advertised vacancies, not completed hires, layoffs, or retraining outcomes. A separate August 2026 New York Fed synthesis, “AI’s Impact on Labor and Hiring,” reports regional business survey results. Its table shows the share of service firms reporting AI use rising from 25 percent in 2024 to 40 percent in 2025, and manufacturers from 16 to 26 percent; both groups also reported expected use over the next six months. The synthesis says firms reported few AI-driven layoffs so far, while describing changing skills and hiring plans and an intention to retrain workers. Survey responses and intentions are not verified retraining counts or a national displacement measure, nor are they outcomes for the 2025 paper’s participants. A counterpoint comes from the May 2026 abstract “Did US Worker Retraining Reduce Participant Automation Exposure?”, hosted by Google DeepMind and published on arXiv. It analyzes more than 23 million WIOA participation records from 2017 through 2023 and defines its Retrainability Index through wage recovery and changes in Routine Task Intensity. The abstract reports that programs rarely moved workers into less routine-intensive work, many returned to their prior field, and wage gains may reflect catch-up mean reversion; outcomes varied with prior skills, location, and the local economy. It also reports an association between employer-led programs, especially apprenticeships, and higher incidence of success. This qualifies optimistic readings of training: earnings recovery need not mean occupational transformation. But this study uses a different window, record set, index, and outcome from Staff Report 1165. This abstract cannot establish all methodological details; its findings are neither replication nor contradiction of the 2025 estimate. Taken together, these sources favor keeping three questions separate: what tasks are exposed, what local employers advertise or report, and what happened to participants in a defined training program. For a career decision, inspect your changing tasks and a modest sample of local postings, then ask what requirements recur and whether an adjacent move can use existing skills. Treat postings as stated demand, not proof of hiring; survey plans as plans; and program averages as group results. The newer signals do not justify a personal job-loss percentage, a universal move away from exposed work, or automatic investment in AI specialization.

Sources: Do Job Postings Show Early Labor-Market Effects of AI?; AI’s Impact on Labor and Hiring; Did US Worker Retraining Reduce Participant Automation Exposure?

How should you compare a course, an adjacent move, and a larger pivot?

Compare each path against the same destination and constraints. A workflow upgrade asks whether you can handle one changing task better in your current role. An adjacent move asks whether your domain knowledge transfers to a nearby role with a different task mix. A larger pivot asks whether a verified gap blocks both options and justifies rebuilding more of your skill set. A course, credential, project, or self-study plan can test or close a gap; completion alone does not establish hiring value. The New York Fed’s How Retrainable Are AI-Exposed Workers? gives a reason to compare paths instead of assuming the most AI-specific one is best. Among trainees from high-exposure occupations, those targeting high-AI-exposure occupations had estimated quarterly earnings returns of roughly $1,040, about 29 percent below peers pursuing more general training. Both groups had positive estimated returns relative to matched job-search-assistance controls. This observational comparison concerns program participants’ target occupations, not a trial of particular courses or degrees. Treat it as a prompt to investigate the target work, not as a ranking of credentials. Begin with the work and destination. List changing tasks, decisions you remain accountable for, and work you want to do more of. Then inspect current postings where you can realistically work. Note repeated requirements, such as a data skill, domain experience, credential, or output review. Postings show advertised requirements, not hires or proof that a listed credential causes a transition. The New York Fed’s Do Job Postings Show Early Labor-Market Effects of AI? analyzes vacancies with an exposure measure informed by observed AI use; postings are a labor-demand signal, not worker outcomes. Use them to form questions, not forecasts. For each option, compare task overlap, prerequisites, feedback or supervised practice, time, full cost, likely salary floor, location, and what your health and family can accommodate. A workflow upgrade might map one reporting task, test an approved tool on suitable material, and record where checking remains necessary. An adjacent move might need a portfolio sample or targeted course that makes existing expertise clear. A larger pivot may require foundational study, formal prerequisites, sustained practice, and a period when earnings or schedule could change. Include tuition, materials, travel, unpaid study time, and delayed plans, not course length alone. Choose learning by intended outcome. To use AI in an existing field, build durable foundations: problem framing, data literacy, evaluation, domain knowledge, verification, and basic automation. These help judge outputs across products. A vendor interface may help with a named task, but a specific menu is less durable. Building AI-enabled software, software practice, and machine-learning engineering or research require different depth and prerequisites. A short course can introduce concepts or support a project; it should not be mistaken for the preparation those technical goals may require. A certificate can structure learning and signal effort, but its value depends on whether target employers request or recognize it and whether you can demonstrate the work. A degree offers broader, longer study and may be required for some roles, with greater time and cost. Self-study can be flexible but requires you to arrange sequence, feedback, and proof of skill. A project demonstrates a capability, though it may not replace a required credential or supervised experience. Check provider curriculum, prerequisites, fees, and duration; assess hiring claims against independent evidence and local requirements. None guarantees a job. A substantial AI-specific qualification makes sense when a target role is real, local demand repeatedly names its requirements, you know which prerequisite you lack, and the time and cost fit your salary floor and other constraints. If those conditions remain unclear, gather evidence with a bounded project or low-cost task test first. Write three options—stay and redesign a task, move adjacent, or make a larger change—and for each note the evidence, the gap, the next affordable test, and what would make you continue or stop. This cannot remove labor-market uncertainty, but it keeps an occupation-level exposure label from making the decision for you.

Sources: How Retrainable Are AI-Exposed Workers?; Do Job Postings Show Early Labor-Market Effects of AI?

What is one proportionate next move?

Pick one recurring task, list its steps, and mark where review or approval enters. If workplace rules permit an AI tool and the material is suitable, test it on a low-risk part. Compare the output with a trusted standard, record errors or rework, and note who remains accountable. The New York Fed’s Staff Report 1165 studies historical training outcomes for program participants; it cannot tell you whether this workflow test will improve your prospects. It does support evaluating training choices rather than assuming an exposed occupation requires an immediate exit. Then inspect a few postings in your location for a nearby role. Note recurring requirements, prerequisites, salary when disclosed, and location or schedule constraints. “Do Job Postings Show Early Labor-Market Effects of AI?” measures advertised vacancies, which can inform a demand check but cannot establish hires, layoffs, or your likely outcome. Compare redesigning your role, investigating an adjacent one, and making a larger change if a specific gap warrants it. Weigh each against your income floor, time, cost, health, and family needs. The free task checker organizes task-level change-pressure signals; these are not a validated displacement probability. A roadmap can help compare scenarios and constraints, without guaranteeing an outcome. Start with the smallest test. Expand training only if it reveals a specific gap and the route fits your circumstances.

Sources: How Retrainable Are AI-Exposed Workers?; Do Job Postings Show Early Labor-Market Effects of AI?

Questions readers ask