JobRiskAI

What to do if your exposure score is high

Short answer

Be sceptical of confident advice, including the encouraging kind. The largest United States retraining evaluations found no earnings benefit from training itself within their follow-up windows, while structured job-search help produced measurable gains quickly and a handful of long, employer-linked sector programmes produced large ones over years. Meanwhile the widely reported shortfall in employment of 22 to 25 year olds in AI-exposed work comes from payroll records whose own authors describe them as descriptive rather than causal, and the gap is far smaller in nationally representative data; the mechanism they identify is slower hiring, not layoffs. In a large field study of a generative AI assistant, the measured productivity gains landed mostly on less experienced workers, while Danish evidence from eleven exposed occupations found none. Taken together: use the tools on your own exposed activities, treat the search as work in its own right, and choose long employer-linked retraining over short general courses.

First, be precise about what your score says

A high score means a meaningful share of your occupation's activities overlap with what AI gets used for. It is not a probability, it has no date attached, and it is not evidence that your employer plans anything. If you have not already, exposure, automation and replacement is worth five minutes before you make any decision on the strength of a number.

What the retraining evidence actually shows

The United States Department of Labor's evaluation of Trade Adjustment Assistance, the flagship programme for workers displaced by structural change, compared participants against a matched comparison group and found participants earned $37,133 less over four years,1 with the programme costed at a net social loss of $53,802 per participant, or $27,494 against an alternative comparison group of unemployment insurance exhaustees.2

Most of the gap accumulates while participants are in training rather than working: by the fourth year the difference had narrowed to about $3,300 a year and there was no significant difference in weeks worked.1 About 66% of TAA-funded trainees completed their programme, and among trainees who were employed in that fourth year, only about 37% were working in the occupation they trained for.3

The random-assignment evaluation of the current United States workforce system found no significant difference in employment or earnings between people offered training and people offered only lighter-touch services over a thirty-month follow-up. Its own evaluators call that evidence not conclusive: only a minority of the group offered training took it up, and many in the comparison group obtained training elsewhere, which narrows the contrast the trial can detect.4

The broadest international meta-analysis finds average effects near zero immediately after a programme, rising to modest positive effects after two to three years. Those medium and longer-run gains are concentrated in human-capital programmes, which is to say training, while job-search assistance shows its advantage mainly in the short run.5

So the honest summary is not "training does not work". It is that training pays slowly and unevenly, that short follow-ups make it look worse than it is, and that generic retraining chosen after displacement has a weak record compared with the promises made for it.

What worked fastest, and what worked biggest

Structured job-search help, which is cheap and quick. In the workforce-system trial, people who received staff-assisted job search and case management earned $3,300 to $7,100 more, 7% to 20%, over thirty months than those left with self-service only, the size depending on which data source is used.4

Long, employer-linked sector programmes, which are slow and large. A randomised trial of Project QUEST in San Antonio, following 410 low-income adults entering healthcare training at community colleges with sustained financial and personal support, found annual earnings gains above $5,000 by year nine, and gains of 15% to 20% in years nine to fourteen. Those gains were statistically significant or close to it in years nine to twelve, and fell short of significance in years thirteen and fourteen, which the evaluators suggest may reflect earnings growing more dispersed over time.5

The practical translation: treat the search itself as real work, and if you retrain, choose something long, specific and connected to employers rather than short and general.

Who actually gains from using the tools

A firm-level field study of 5,172 customer support agents using a generative AI assistant found productivity up 15% on average. The distribution matters more than the average: gains ran around 30% for less experienced and lower-skilled workers, while the most experienced saw small speed gains together with small declines in output quality.6 Against that, the Danish evidence covered in exposure, automation and replacement found no earnings or hours effect at all, which limits how far a single-firm result generalises.7

If you are early in your career, these tools appear to be a larger multiplier for you than for your senior colleagues. If you are experienced, your advantage is judgment and review rather than raw speed.

Where the pressure is actually appearing

Using payroll records covering millions of workers, researchers at Stanford report that by June 2026 employment of 22 to 25 year olds in AI-exposed occupations sat 19% below where it would have been had it kept pace with less-exposed peers, with no comparable gap for experienced workers.8

The authors' first stated finding is that they see "no evidence of widespread, economy-wide job displacement". They state plainly that these are "descriptive facts" and not a causal estimate of AI. The mechanism they identify is reduced hiring of young workers rather than increased separations. And over the window where a nationally representative comparison is possible, the gap is about six times smaller in American Community Survey data than in their payroll sample.8 An earlier and much-quoted version of this work headlined 13%; in the August 2026 revision the authors lead with the simpler descriptive measure instead, and note that holding the original specification fixed on current data, its estimate for the most exposed quintile is no longer statistically significant.8

Independent data points the same way without settling the cause. On Indeed, software development postings sat about 27.5% below their pre-pandemic level in mid-2026 while postings overall were back to roughly February 2020 levels. That category had been recovering since early 2025, but 71% of the increase came from senior roles and 37% from jobs with AI in the title.9 A recovery that skips the bottom rung is consistent with the payroll finding.

The honest complication: the collapse in tech postings began in mid-2022, several months before ChatGPT was public.9 Interest rates and post-pandemic over-hiring were already doing work. Anyone who tells you the entry-level squeeze is purely an AI story is going beyond the evidence, and so is anyone who tells you AI has nothing to do with it.

What this means for you

Use the tools on your own most-exposed activities now, because the measured gains are real and they are largest for the less experienced. Shift your visible contribution toward judgment, review and accountability: this one is an inference rather than a measured finding, but it follows from the pattern that directing activities are the ones that barely register in Microsoft's classified usage shares.10 If you are considering retraining, choose employer-linked and specific over short and general, and treat the job search as work in its own right. And if you are early in your career, assume the entry rung is narrower than it was and compete on demonstrated output rather than credentials.

One measurable piece of leverage: in Lightcast's 2025 analysis, job postings that asked for at least one AI skill advertised median salaries about 28% higher, roughly $18,000 a year, than postings that did not.11 Treat that as a signal about what employers currently advertise rather than a promise about your own pay, because postings are not vacancies and the firms hiring for AI skills differ from those that are not.

Your own occupation page turns this into specifics: which of your activities carry the overlap, which do not, and a plan built around that split. Start from the calculator, or compare your role against an adjacent one with the comparison tool.

Sources
  1. Peter Z. Schochet, Ronald D'Amico, Jillian Berk, Sarah Dolfin and Nathan Wozny, Estimated Impacts for Participants in the Trade Adjustment Assistance Program Under the 2002 Amendments, Mathematica Policy Research for the U.S. Department of Labor. dol.gov evaluation studies. Quoted: total earnings over quarters 1 to 16 of $42,939 for participants against $80,072 for the matched comparison group, an estimated impact of minus $37,133; and a year-four annual earnings gap of about $3,273 with a statistically insignificant difference of two weeks worked. These are non-experimental matched-comparison estimates.
  2. Sarah Dolfin and Peter Z. Schochet, The Benefits and Costs of the Trade Adjustment Assistance (TAA) Program Under the 2002 Amendments, Mathematica Policy Research for the U.S. Department of Labor, December 2012 (ETAOP 2013-09). Quoted: an estimated net social cost of $53,802 per participant against the matched comparison group, falling to $27,494 against an alternative comparison group of unemployment insurance exhaustees. Benefit-cost estimates of this kind rest on assumptions about how long earnings differences persist beyond the observation window.
  3. Jillian Berk, Understanding the Employment Outcomes of Trainees in the Trade Adjustment Assistance (TAA) Program Under the 2002 Amendments, Mathematica Policy Research for the U.S. Department of Labor, December 2012 (ETAOP 2013-11). Quoted: 65.9% of TAA-funded trainees completing a training program, and, among TAA trainees who were working in the final year of follow-up, 37% employed in the occupations for which they trained. The report notes that a match is counted only at the two-digit occupational code level, so the 37% may overstate or understate the true rate.
  4. Kenneth Fortson, Dana Rotz, Paul Burkander, Annalisa Mastri, Peter Schochet, Linda Rosenberg, Sheena McConnell and Ronald D'Amico, Providing Public Workforce Services to Job Seekers: 30-Month Impact Findings on the WIA Adult and Dislocated Worker Programs, Mathematica Policy Research for the U.S. Department of Labor, 2017. mathematica.org. Quoted: no significant employment or earnings differences between the full-WIA and core-and-intensive groups over thirty months, with the evaluators noting the evidence on training is not conclusive because of limited take-up and substitution; and core-and-intensive customers earning $3,300 to $7,100 (7% to 20%) more than core-only customers over the same period depending on the data source.
  5. David Card, Jochen Kluve and Andrea Weber, What Works? A Meta Analysis of Recent Active Labor Market Program Evaluations, Journal of the European Economic Association, 2018 (average effect sizes of 0.04 standard deviations short run, 0.12 medium run and 0.19 longer run, with medium and longer-run gains concentrated in human-capital programmes); and Anne Roder and Mark Elliott, Fourteen Year Gains: Project QUEST's Remarkable Impact, Economic Mobility Corporation, October 2024 (full report, PDF), with the earlier Nine Year Gains, 2019, and the review by Social Programs That Work. 410 low-income adults randomly assigned in San Antonio entering healthcare training with sustained support. The year-nine impact of $5,239 is the 2019 report's figure in current dollars; the 2024 report puts that same year at $6,574 in constant 2022 dollars. The gains of 15% to 20% across years nine to fourteen, averaging about $6,000 a year in constant 2022 dollars, are Social Programs That Work's calculation from the 2024 report's own earnings table rather than the report's wording; the report describes increases of 15 to 26 percent per year across the nine of the last eleven years in which the earnings gap exceeded $4,500, a wider window. Impacts are statistically significant in years nine and ten, near-significant in years eleven and twelve, and short of significance in years thirteen and fourteen.
  6. Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond, Generative AI at Work, Quarterly Journal of Economics 140(2), 2025, pp. 889-942. QJE. Quoted from the published version: 15% average increase in issues resolved per hour across 5,172 agents; roughly 30% gains for less experienced and lower-skilled workers; small speed gains with small quality declines for the most experienced. The widely repeated 14% and 34% pair comes from the 2023 working paper, not the published article. The authors state the study is not designed to measure aggregate employment or wage effects, and bound it to the medium run within a single firm.
  7. Anders Humlum and Emilie Vestergaard, National Bureau of Economic Research Working Paper 33777, 2025, revised 2026. nber.org. About 25,000 Danish workers across 7,000 workplaces in eleven highly exposed occupations, with null effects on earnings and hours two years after ChatGPT's launch.
  8. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Stanford Digital Economy Lab, August 2025, revised August 2026. digitaleconomy.stanford.edu. Quoted from the August 2026 version: no evidence of widespread economy-wide displacement; the 19% kept-pace shortfall for workers aged 22 to 25 as of June 2026; adjustment through reduced hiring rather than separations; the explicit statement that these are descriptive facts and not a causal estimate; the roughly six-fold smaller gap in American Community Survey data than in the ADP payroll sample over the comparable window; and the superseded 13% figure, whose original specification the authors report is no longer statistically significant for the most exposed quintile on current data. ADP provided the data and supports the Stanford Digital Economy Lab through its Corporate Affiliate Program and a separate unrelated sponsored project, and has the right to review the paper to prevent disclosure of confidential information.
  9. Guillermo Gallacher, AI and Job Postings: From Destruction to Creation?, Indeed Hiring Lab, 8 July 2026. hiringlab.indeed.com. Software development postings about 27.5% below their pre-pandemic level against overall postings essentially at February 2020 levels; 71% of the increase in software development postings between May 2025 and May 2026 from senior roles and 37% from postings with AI in the title; and the decline in AI-exposed vacancies beginning before ChatGPT's release in late 2022, which Indeed attributes to other researchers rather than to its own analysis. Indeed states the relationship between AI exposure and postings changes is uncontrolled.
  10. Microsoft Research, Working with AI published result files, CC BY 4.0. github.com/microsoft/working-with-ai. All six activities whose titles begin "direct" score zero on the user-goal series, computed from iwa_metrics.csv on 22 August 2026. A zero means the activity fell below a usage share of about 0.05%, not that AI never did it. These shares come from a pipeline that maps conversations onto O*NET work activities with a large language model; the paper reports agreement with human annotators at Cohen's kappa of 0.35 to 0.44 for user goals and 0.34 to 0.53 for AI actions, and Microsoft states that all metrics should be read with caution. The recommendation to shift toward judgment and review is an inference from that pattern, not a measured career outcome.
  11. Lightcast, Beyond the Buzz: Developing the AI Skills Employers Actually Need, 2025. lightcast.io. In Lightcast's 2025 analysis, postings requesting at least one AI skill advertised median salaries about 28% higher, roughly $18,000 per advertised annual salary, than comparable postings without. Lightcast reissues this research periodically, so a later edition may report a different premium. The figure is drawn from Lightcast's global postings collection rather than a United States-only sample. Advertised salaries in job postings are not the same as vacancies or as paid wages, and Lightcast's skill classification is produced by a language model.

Every figure above is quoted with the caveat its own source states. Last verified 2026-08-22. Scores and bands used on this site are documented on the methodology page.