The New York Fed analyzed roughly 109,000 WIOA/WIA training spells from 2012–2023 in its restricted sample; the report’s larger 1.6 million figure counts participation spells, and only 22% had occupation codes. Using nearest-neighbor matching, the researchers compared eligible trainees with Wagner-Peyser job-search-service recipients and estimated short-run earnings differences. For trainees from high-AI-exposure occupations, the reported quarterly difference was about $1,470. In the destination comparison, high-exposure trainees targeting high-AI-intensity occupations had a positive estimated return of about $1,039, roughly 29% below the return for peers whose training could lead to any occupation. These are matched observational results among participants with usable records, not randomized effects or a head-to-head test of changing fields versus taking an AI course; they do not predict an individual’s job-loss risk or long-term earnings. A practical next step is to compare an adjacent role with a deeper technical path against actual requirements, local demand, pay, time, and cost.
What did the retraining study actually compare?
The New York Fed report’s headline dataset contains more than 1.6 million U.S. Workforce Investment and Opportunity Act (WIA) and Workforce Innovation and Opportunity Act (WIOA) participation spells from 2012 through 2023. That is not the size of the restricted earnings analysis: it contains about 109,000 training spells, and only 22 percent had occupation codes. For its matched analysis, researchers linked program records to occupational AI-exposure measures and used nearest-neighbor matching to compare eligible training participants with Wagner-Peyser recipients who received job-search assistance. Among trainees from high-exposure occupations, the estimated earnings difference was about $1,470 per quarter. This is a short-run matched observational estimate, not a randomized effect or a forecast for an individual; the restriction to participants with usable earnings and occupation data limits who the result describes. [New York Fed study](https://www.newyorkfed.org/research/staff_reports/sr1165)
The destination comparison adds a separate result. Among trainees from highly exposed occupations, those targeting high-AI-intensity occupations had a positive estimated return of about $1,039, roughly 29 percent below the return for peers whose training could lead to any occupation. The report’s comparison is between those destination groups, with matched job-search-service recipients as the reference—not a randomized head-to-head trial of changing fields versus taking an AI course. “High AI intensity” is an occupation-level classification; it does not mean these workers all entered machine-learning engineering, took the same course, or mastered a specific tool. The broader destination group also does not mean every trainee changed fields; some may have used general training to stay in related work.
The study therefore distinguishes prior-occupation exposure from intended destination intensity. It does not identify why the estimated returns differed, include all private or employer training, or establish long-term career paths. Positive earnings results were driven largely by later years when labor markets were tighter. The final matched analysis also excludes people without the continuous observed earnings and occupation data required for inclusion. The fair reading is encouraging but bounded: among the studied public-program participants with usable records, training was associated with positive short-run earnings differences for workers from high-exposure occupations, while the high-intensity destination group had a smaller relative return. Neither result estimates displacement probability or proves which route will pay for an individual worker.
Which move fits the work and constraints I already have?
Compare three actual work paths before comparing labels. An upgrade keeps you in your field while improving a recurring task: for example, using an approved tool to draft a first pass, then checking source records, exceptions, and decisions for which you remain accountable. An adjacent move changes the task mix but carries forward domain knowledge. Imagine a billing analyst learning data-quality workflows or a project coordinator shifting toward process testing; those are illustrations, not reported study outcomes. A deeper technical transition aims at building, evaluating, or maintaining AI systems and usually asks for more computing, statistics, or engineering preparation. Practical AI use inside a field is not the same goal as becoming an AI developer or researcher.
The U.S. Bureau of Labor Statistics’ 2025–35 projections show why a broad “AI jobs are growing” headline is too thin for choosing. Its current table projects 34.6 percent employment growth for data scientists and 10.2 percent for software developers, and lists a bachelor’s degree as the typical entry education for each. These are national occupation projections, not local vacancies, entry odds, or evidence that a particular credential causes a hire. The table helps identify roles to investigate; it does not establish that a career changer can meet the requirements quickly or at an acceptable cost. [BLS occupational projections](https://www.bls.gov/emp/tables/occupational-projections-and-characteristics.htm)
Choose learning by the gap in the target work. A short course can teach a bounded concept or tool workflow, but completion alone does not demonstrate job readiness. A certificate may offer structure or a signal if employers in your target market request it; check actual postings before paying. A project can show that you applied data handling, evaluation, or workflow design to a real problem, especially when you can explain errors and human checks. Self-study is flexible but offers less feedback. A degree takes more time and cost, with broader depth and formal recognition; use it when the target role’s prerequisites or your learning goal justify that commitment. For machine-learning research or engineering, a survey course is not a substitute for the math, programming, systems practice, and deeper study those roles may require. The right sequence depends on whether your goal is AI use in your current field, a technical build role, or research.
For each route, write down the target task bundle and compare prerequisites, credible local hiring signals, salary floor, commute or remote limits, training time and cost, and family or health constraints. If the evidence is thin, a modest course or supervised project can test interest and aptitude before a degree or expensive program. That is a decision rule inferred from the study and role evidence, not a ranking its researchers tested.
Sources: Occupational projections and worker characteristics
What is a proportionate next step before retraining?
Start with tasks rather than a whole-job label. List work that is digital and repeatable, such as sorting routine requests or preparing a standard first draft. Separately mark tasks that depend on exception judgment, customer context, physical conditions, or accountability for a consequential decision. A tool may be capable of contributing to a task without your employer having adopted it; adoption may change workflow without reducing headcount; and labor demand may move for reasons beyond AI. Those are distinct signals. Exposure is a reason to investigate, not an estimate of redundancy.
Then compare one redesign path, one adjacent destination, and one larger technical transition if it matches your goal. For each, find several current local or genuinely accessible postings and note recurring duties and explicit prerequisites. Check whether the expected pay could meet your floor and whether the location, schedule, and training burden fit your life. National BLS projections and global employer plans are useful background, but they cannot answer those local questions. If relevant postings mostly request domain experience plus data handling, the next learning step may be a focused project. If the roles consistently require deeper programming and quantitative foundations, plan that gap honestly. If no accessible roles meet your pay or location limits, that is evidence to revise the destination before buying training.
The evidence could support deeper specialization when a concrete technical destination repeatedly asks for skills you want to build, the prerequisites are attainable, and local demand and training support justify the investment. An adjacent move is a more proportionate first test when it preserves valuable domain knowledge and the technical destination remains vague. Neither route makes a job immune to change. This conclusion would shift if local employers offered funded technical training, if the target role’s requirements changed, or if nearby work no longer met your constraints.
This week, choose one recurring task that takes meaningful time, record where errors would need human review, and compare it with one target role’s posted duties and requirements. The free [AI task checker](/ai-job-risk-checker) can organize task-level change-pressure signals and first actions; it does not predict redundancy or select a career. If you need to compare a stay-and-redesign option, an adjacent path, and a larger change against your own pay, location, time, and family constraints, the [career roadmap](/career-roadmap) is designed for that planning decision and does not guarantee employment or income. To make the research practical, ask your manager, mentor, or training adviser: “Could we test this task on a supervised pilot, and what evidence would show that I need more training before taking responsibility for it?”
Sources: How Retrainable Are AI-Exposed Workers?; Occupational projections and worker characteristics; Appendix — The Future of Jobs Report 2025; 4. Workforce strategies — The Future of Jobs Report 2025
Questions readers ask
Does the New York Fed study show that AI-exposed workers should avoid AI careers?
No. It found positive short-term earnings differences for trainees from high-exposure occupations, including those targeting high-AI-intensity work. That target group had lower relative returns than peers with broader destinations. The study did not compare AI courses with field changes or establish individual outcomes.
Should I get a degree, certificate, or take a course before changing fields?
First identify a specific target role and inspect current accessible postings for repeated prerequisites. Use a course for a defined gap, a project or supervised task to demonstrate applied work, and a certificate when the target employers value it. A degree makes sense when its depth or formal requirement justifies the time and cost.
Sources and notes
- How Retrainable Are AI-Exposed Workers?
The report’s dataset includes more than 1.6 million WIA/WIOA participation spells from 2012–2023; its restricted matched analysis contains about 109,000 training spells, and only 22% had occupation codes. The authors use nearest-neighbor matching to compare eligible trainees with Wagner-Peyser job-search-assistance recipients. For trainees from high-AI-exposure occupations, the estimated short-run quarterly earnings difference was about $1,470. Within the destination comparison, those targeting high-AI-intensity occupations had a positive estimated return around $1,039, about 29% lower than peers whose training could lead to any occupation. This observational analysis is restricted to participants with usable records; it is not a randomized estimate, a clean new-field versus AI-course test, an individual guarantee, or evidence of long-term outcomes.
- Occupational projections and worker characteristics
BLS Table 1.2 lists 2025–2035 national employment growth projections of 34.6% for data scientists and 10.2% for software developers, with a bachelor's degree as typical entry education for each. These are national projections and occupational characteristics, not local vacancy counts, individual entry odds, or causal evidence about credentials.
- 4. Workforce strategies — The Future of Jobs Report 2025
The report states that 77% of surveyed employers plan to reskill or upskill existing workers to work more effectively alongside AI by 2030. This is a stated plan in the report, not an observed outcome or local hiring forecast.
- Appendix — The Future of Jobs Report 2025
The methodology describes a late-2024 survey of senior leaders at global employers: 1,043 company responses representing over 14.1 million employees across 22 industry clusters and 55 economies. The survey targeted large companies, with a threshold of 500 employees, and explicitly excludes small enterprises and the informal sector from scope. Its reported percentages represent respondent organizations' self-reported strategies and expectations, not population-wide measures or forecasts for a specific local labor market.
Apply this to your own work
See the whole job market at once.
Explore which occupations AI may reshape, then turn the signal into a practical response.
Explore the job map