AI is now the most-cited reason for layoffs in America. It is also invisible in the national employment data. Both are true, and the reason is scale.
- Employers are naming AI in job cuts. It has been the leading reason given in announced US cuts for months running, though an announcement is not a verified cause.
- The national data barely moves. The cuts are a tiny share of the workforce, and unemployment rose slightly less in the most AI-exposed jobs than in the least exposed.
- The clearest effect is on first jobs. Young workers in the most exposed occupations are being hired less rather than laid off, though the researchers say AI may not be the cause.
- “AI-proof” lists are opinion. Resistance scores publish no method, and the old paper they lean on estimated what could be automated, not what would be.
- Serious measures disagree. Two methods put some jobs on opposite sides, so judge the tasks in your own week rather than a job title.
↓ Jump straight to the list of high-risk and low-risk jobs — grouped by how good the evidence is.
“What jobs are safe from AI” is one of the most-asked questions on the internet and one of the worst-served. The answers are dominated by listicles with confident numbers and no method. This page does the opposite: it starts with what has actually been measured, says plainly how little that is, and only then talks about what looks durable.
1 · Two true things that sound contradictory
Almost every piece written about this picks one of two stories. Either AI is gutting employment, or it is a bubble and nothing is happening. Telling those apart is a claim-checking problem before it is an economics one. The evidence supports both at once, and understanding why is most of the answer.
People are losing jobs to AI, by name. Challenger, Gray & Christmas track the reasons employers give for announced job cuts. Since they began recording AI as a category in 2023, employers have cited it in 184,538 announced cuts — 112,713 of those in 2026 alone through July, roughly 24% of all cuts this year. It has been the single leading stated reason for five consecutive months.Challenger, Gray & Christmas, monthly Job Cut Announcement Report, 6 Aug 2026, read at source 9 Sep 2026 — “10,970 announced during the month, or 33%”, “112,713 job cut announcements, approximately 24% of all cuts”, “184,538 job cut announcements” since 2023. July 10,970 AI-attributed cuts (33% of the month), 2026 YTD 112,713 of 477,033 total, cumulative since 2023 184,538.
And the national data shows nothing. Yale’s Budget Lab tracks the occupational mix — what jobs Americans actually hold — on the reasoning that mass displacement would show up as that mix changing faster than usual. It has not; the mild acceleration predates ChatGPT.The Budget Lab at Yale, Tracking the Impact of AI on the Labor Market, updated through early 2026, read at source 9 Sep 2026: “shifts in the occupational mix were well on their way during 2021, before the release of generative AI, and more recent changes do not seem any more pronounced”. More pointedly: since 2022 unemployment among the most AI-exposed workers rose 0.77 percentage points, while among the least exposed it rose 0.85.Stanford Institute for Economic Policy Research, What is really happening to jobs? Separating AI hype from reality, 2026, read at source 9 Sep 2026 — it finds no economy-wide AI jobs apocalypse while reporting “a notable decline in employment among early-career workers in AI-exposed occupations”; using Census/IPUMS-CPS data 2022–2026.
184,538 announced cuts across three years, against US employment of about 170 million, is roughly a tenth of one percent. That is a large number of people and a rounding error in a national statistic. The unemployment rate cannot see it; the people in it certainly can.Reasoning, September 2026 — arithmetic: 184,538 ÷ 170.3 million ≈ 0.11%, using the Challenger total above and the BLS 2025 employment base of 170.3 million cited in the projections section below.
Anyone telling you AI has taken no jobs is reading the aggregate and ignoring the individuals. Anyone telling you it is causing mass unemployment is doing the reverse.
Two caveats, in opposite directions. Challenger counts announcements and the reason the employer gave — not verified separations, and not an independent finding of cause. Several analysts have documented firms attributing routine cuts to AI because it reads better to investors, which inflates the figure. But the reverse also holds: a role quietly never backfilled is a job lost to AI that appears in no announcement at all.
Hold both facts at once: AI is costing real people their jobs, and the effect is still too small to show in national figures.
2 · The one real effect is the bottom rung
There is a genuine, measured exception, and it is narrow: people trying to get their first job.
Employment for 22–25 year olds in the most AI-exposed occupations — customer service representatives, secretaries and administrative assistants, first-line retail supervisors — fell about 13% since 2022. The mechanism matters: the Dallas Fed found it is driven by fewer people moving from outside the workforce into employment, not by layoffs or by the unemployed failing to find work.Federal Reserve Bank of Dallas, Young workers’ employment drops in occupations with high AI exposure, Jan 2026, read at source 9 Sep 2026 — the fall “is not due to layoffs but to a low job finding rate for young workers entering the labor force”; on work by Brynjolfsson and colleagues at the Stanford Digital Economy Lab.
Two honest qualifications, both from the researchers themselves. If that entire decline became unemployment it would add 0.1 points to the national rate. And the pattern may not be causal at all — it coincides with interest rate rises and the unwinding of pandemic over-hiring, which are hard to separate from AI.
An effect that is small today and concentrated on new entrants is exactly what a slow structural change looks like early. In the Dallas Fed’s reading it is not layoffs; the door is simply opened less often. It takes years to show up in aggregate statistics and it lands entirely on people with no track record to fall back on.
This site covers that specific argument in more depth in AI and Work: what the evidence actually shows, including the evidence pointing the other way.
If you are starting out, the measured risk is fewer openings rather than layoffs, so anything that stands in for a track record counts for more.
3 · Why the “AI-proof jobs” lists are not evidence
Search the question and you will find ranked lists with numbers attached — mental health work scored 95 out of 100 for AI resistance, skilled trades 91, and so on. Those scores are not measurements. In the cases traced for this page they originate on the marketing blogs of companies selling AI products, with no stated method.
The better-sourced lists have a subtler problem. A widely syndicated 2026 feature titled 15 Fast-Growing Jobs AI Can’t Touch is a repackaging of a report by a résumé-writing company. Its method is to filter official labour statistics for jobs paying over $49,500 and growing more than 10%, then apply judgement about which of those “resist AI”.Resume Genius, 15 Fastest-Growing AI-Proof Jobs in 2026, read at source 9 Sep 2026 — its filter is a “median annual salary of $49,500, projected job growth exceeding 10% over 2024–2034, and demonstrated resilience to AI displacement per the Frey and Osborne framework”; syndicated by Forbes, 16 Aug 2026 (Forbes blocks automated and headless access, so that page was not read here). In the current BLS round (2025–35), nurse practitioners are the fastest-growing occupation, at 41%, ahead of solar photovoltaic installers (37%) and wind turbine service technicians (30%), read at source 17 Sep 2026.
The growth half is solid — it is government data. The “AI can’t touch” half is unmeasured opinion. And the framework it cites for the resistance judgement is Frey and Osborne’s 2013 paper. That paper estimated 47% of US jobs were susceptible to computerisation — technically automatable at some point, which its authors were explicit was not a forecast that they would be. Thirteen years on it is being cited as though it were one, to tell people which careers are safe.Frey & Osborne, The Future of Employment: How Susceptible Are Jobs to Computerisation?, Oxford Martin School, 2013, read at source 9 Sep 2026. The paper estimates technical susceptibility, and names complex perception, creativity and social intelligence as barriers to actual automation.
Trust the growth projections, discount “AI can’t touch this” labels, and ignore resistance scores that publish no method.
4 · When someone measured it properly, the answers inverted
In May 2026 two researchers tried a different approach. Instead of asking whether a job looks automatable, they scored all 17,951 tasks in the US occupational database for whether an AI system could plausibly be trained to do them — whether the task can be framed as something with a clear objective a model can be rewarded for completing.Tomei & Klein Teeselink, What Jobs Can AI Learn? Measuring Exposure by Reinforcement Learning, arXiv:2605.02598, May 2026 — an RL Feasibility Index across all 17,951 O*NET tasks.
Their results disagree with the conventional exposure rankings in both directions. Physicians, musicians and natural sciences managers score low on what AI can actually be trained to do, while conventional indices rate them highly exposed. Power plant operators, railroad conductors and aircraft cargo handling supervisors score the reverse — routinely listed as safe manual work, but highly trainable.
The lesson is not that this index is the truth. It is that two serious attempts to rank the same jobs produce opposite answers for some of them, which is a reason to hold any single ranking loosely.
5 · What actually looks durable, and why
Strip out the rankings and four properties do the real work. None of them is a guarantee; each is a reason a task has been hard to hand over so far.
6 · The list, sorted by how much we actually know
Here is the part everyone comes for. It is grouped by evidence quality rather than ranked, because a job where employment has measurably fallen and a job somebody thinks feels automatable are not the same kind of claim, and putting them in one ranked column is how the bad lists get made.
Group A is what has already happened. Group B is where the cuts are being announced. Group C is where the evidence is contested. Group D is what has held up so far. Nothing here is a prediction about your job.
A · Occupations where employment has actually fallen
The Bureau of Labor Statistics identified 18 occupations as AI-exposed. Between May 2024 and May 2025, employment across that group fell 0.2% while total US employment rose 0.8%. Strip out medical secretaries — which grew for unrelated reasons — and the rest fell 1.6%.US Bureau of Labor Statistics, Occupational Employment and Wage Statistics, table 1 of the May 2024 and May 2025 releases, read at source 17 Sep 2026: customer service representatives 2,725,930 then 2,595,750, a fall of 130,180 (4.8%); all occupations 154,187,380 then 155,495,730, a rise of 0.8%. The 0.2% fall for the 18-occupation group was not recomputed here. The medical-secretary adjustment is Bloomberg’s analysis of the same release, which is paywalled and was not re-read here.
Note what is not on this list: nearly everything the popular articles put on theirs. A 0.2% fall across 18 occupations is a real signal and a small one, and it is the strongest occupation-level evidence this site found.Reasoning, September 2026 — the 0.2% is the group figure from the BLS paragraph above, which this site did not recompute; “strongest” is this site’s judgement across the sources on this page.
B · Where employers say the cuts are
Different question, different answer. This is not measured displacement — it is what companies stated when announcing layoffs, which is self-reported and sometimes flattering.
C · Where the serious measures disagree
These are occupations two credible methods put on opposite sides. That disagreement is information: it means the honest answer is that nobody knows.
D · Least exposed on every measure
About 40% of US workers are in jobs with zero measured AI exposure, and fewer than 10% are in occupations scoring above 0.4 on standard exposure scales. The low end is consistent across methods.Federal Reserve Bank of New York, Liberty Street Economics, Do Job Postings Show Early Labor-Market Effects of AI?, May 2026, read at source 9 Sep 2026: “40 percent of workers are in jobs with zero measured AI exposure” and “less than 10 percent of workers and vacancies are in occupations with an AI exposure of at least 0.4”.
It is not a safety ranking. Group D is low-exposure, not immune — and several of those roles pay poorly, which is its own kind of insecurity.
It is not about your job. Occupations are averages. Two people with the same job title can spend their weeks on completely different tasks, and it is tasks that get automated. The question worth asking is which parts of your week are repeatable with a clear right answer.
7 · Where the jobs actually are
US projections to 2035 have total employment rising just 3.5%, from 170.3 to 176.2 million — 5.9 million jobs over a decade, against 10.9% growth in the decade before. The largest share is in one place: private health care and social assistance, up 9.5% and adding over 2.2 million jobs — more than any other sector, and about 37% of all new jobs. The only sector growing faster in percentage terms is utilities, at 9.8% — and BLS attributes that to electricity demand, including AI’s.US Bureau of Labor Statistics, Employment Projections 2025–35, released 27 Aug 2026, read at source 9 Sep 2026: “total employment is projected to increase from 170.3 million to 176.2 million and grow 3.5 percent, which is slower than the 10.9 percent growth recorded over the 2015-25 decade”, and private healthcare and social assistance “is projected to add the most jobs of any sector, over 2.2 million”, and “This strong growth is expected to account for about 37.0 percent of all new jobs projected to be added through 2035.” An earlier version of this page used the 2024–34 round (3.1%, 170.0→175.2m, health care +8.4%), superseded on 27 Aug 2026.
Nurse practitioners are projected to grow 41% — the fastest-growing occupation in the economy. Home health and personal care aides, already the largest occupation in the country, are projected to add more jobs than any other.US Bureau of Labor Statistics, Employment Projections 2025–35, read at source 17 Sep 2026: “Notably, nurse practitioners is projected to experience the fastest employment growth of all detailed occupations (+41.0 percent), reflecting the central role this occupation plays in providing care under collaborative, team-based models of care.”
The occupation adding the most jobs in the United States is also one of the worst paid. Care work is durable because it is hard to automate and demand is demographic — but durability and pay are separate questions, and a lot of coverage quietly merges them.
One correction to this page’s own record. It previously said that the widely repeated “41%, the fastest-growing job” for nurse practitioners was a drift from a published 40.1% at third-fastest. That was true of the 2024–34 projections. In the 2025–35 round released on 27 August 2026, nurse practitioners are 41% and first. The reporting we criticised became correct before we did. Which is its own lesson about how fast a sourced number goes stale.US Bureau of Labor Statistics, Occupational Outlook Handbook, Fastest Growing Occupations, 2025–35, read at source 9 Sep 2026: nurse practitioners 41%, solar photovoltaic installers 37%, data scientists 35%, wind turbine service technicians 30%.
Alongside those projections BLS launched an official AI exposure dataset — “alongside the 2025-35 projections, BLS is introducing a new data product that provides information about how occupations compare to one another based on their theoretical and observed exposure to artificial intelligence”. Every list in this article predates it. It is the first time the question “which jobs are exposed to AI” has an answer from the agency that counts the jobs.US Bureau of Labor Statistics, AI exposure categories, published 27 Aug 2026, read at source 9 Sep 2026.
Durable does not mean well paid: the occupation adding the most jobs is also one of the worst paid.
8 · The gap between the fear and the effect
ADP surveyed more than 39,000 working adults across 36 markets, in the field 21 July to 4 August 2025. Only 22% strongly agreed their job was safe from elimination. Confidence is lowest among individual contributors and frontline managers, at 18% and 21%. It also varies enormously by country: 38% in Nigeria, 28% in the United States, 25% in the United Kingdom, 5% in Japan.ADP Research, Only 22% of Workers Confident Their Job is Safe from Elimination, Mar 2026, read at source 9 Sep 2026: “only 22% of global workers strongly agreed their job was safe from elimination”, from “over 39,000 working adults across 36 markets”. An earlier version of this page said 38,000 across 34 markets, and attached ADP’s 20/15/10 age figures to job-security confidence — those figures are for whether AI will positively impact the respondent’s job, a different question. Both corrected here.
Set that against the measured picture. Fewer than one worker in four is confident their job is safe from something that has, so far, formally happened to roughly one US worker in a thousand — and which does not move the national unemployment rate at all. The fear is running well ahead of the effect, without the effect being zero.
That gap is worth naming, because it is exploitable. It sells courses, it sells résumé services, and it gives employers cover — several analysts have noted firms attributing routine layoffs to AI, which makes the effect look larger than the data supports.
Be wary of anyone selling to the fear: it is running well ahead of the measured effect, though the effect is not zero.
9 · What to actually do with this
Do not pick a career from a ranking. Look at which tasks in your own week are repeatable with a clear right answer, and check any statistic against its original source.