AI Is Taking Tasks From Your Job Before It Takes the Job

Is the Job Disappearing, or Just Parts of It?

AI is not eliminating whole occupations at the pace the question implies; it is removing individual tasks from many of them, at different rates depending on the task and employer. The research measuring this operates at the level of what a model can already do end to end versus what still needs a person to decide whether the output is right. That is the honest unit for sizing your own risk.

For scale, the Bureau of Labor Statistics projects total US employment growing 3.1 percent across the entire 2024-2034 decade, roughly a third of a percent a year, and it builds that projection from historical relationships without modelling an AI shock as a separate driver ¹. It is a background rate to size a claim against, not evidence for or against one. Use of the tools, meanwhile, is already ordinary: in a nationally representative US survey from late 2024, 23 percent of employed respondents had used generative AI for work at least once in the previous week ². Widespread use and a title disappearing are not the same fact, and treating them as one is most of what makes this question feel more settled than the evidence actually is. Employers are told to run a task-level audit before writing a job posting in the first place: how to find out what AI actually does in a role before writing the job post is that exercise from the hiring side, and reading it shows you what a posting's AI language is supposed to be built on.

Write Out Your Own Week as Tasks

Write out what you actually did last week as a list of discrete tasks rather than as a job title, then mark each one with one of two labels: a model can already do this end to end with light checking, or this needs someone to decide whether the output is right before it goes anywhere. The ratio between those two labels, not your title, is the number that matters.

A worked example makes the exercise concrete. "Drafting a first version of a routine email" usually sits in the first label. "Deciding which of three vendors' numbers to trust in a report" usually sits in the second, even though a model can write sentences about all three. Most real jobs turn out to be a mix, and the mix is what is actually worth tracking over time, not a single verdict about the role.

Do the exercise again in three or six months rather than once. A task that needed a person to check it closely today can move into the first label as tools improve, and a list you keep redoing shows you that movement while it is happening. A single week's version of it, treated as final, tells you nothing about direction, which is the part you can actually respond to.

A close cousin of this exercise runs on the hiring side too, before anyone decides where AI skills genuinely belong on a job posting: which roles actually need AI skills right now, and which are being added for no reason walks through the same audit from that chair. It is worth reading if a posting's AI requirements look nothing like the work the job seems to involve.

Where the Evidence Splits

Two credible bodies of evidence currently disagree, and a reader in 2026 is looking at a live argument rather than a settled finding. Payroll records from one large employer processor, read by researchers at Stanford's Digital Economy Lab, put employment of workers aged 22 to 25 in AI-exposed occupations well below where it would be had it kept pace with less-exposed peers as of June 2026, a gap that shows up in slower hiring rather than in more firing ³⁴.

Those authors call that a descriptive pattern rather than proof that AI caused it, and they report that their own estimate roughly halves once the education level an occupation requires is controlled for ³. That second number is the qualification most often dropped when the finding gets quoted. LinkedIn's own labor market report, published in January 2026 and built on hiring activity across its platform, reaches the opposite headline: entry-level roles have not been disproportionately affected relative to experienced ones over much the same stretch ⁵.

Both are measuring adjacent but different populations by different methods, and neither has resolved the other. The payroll data comes from large employers who share administrative records with one processor and skews toward manufacturing and larger firms. LinkedIn's figure counts members who added a new employer to their profile, so it skews toward white-collar people who keep a profile current, and it is the company's own report on its own platform rather than a peer-reviewed study. Neither sample is a census of the labor market, and a reader is better served by knowing both exist than by picking whichever one confirms what they already expected to hear.

The pattern inside individual tasks helps explain why the aggregate picture is so unsettled. In a 2023 field experiment with 758 consultants at one firm, those given GPT-4 on tasks chosen to sit inside its capability completed more of them, finished faster, and were graded more than 40 percent higher on quality than a control group; on the same experiment's one task chosen to sit outside that capability, the people using the tool were less likely to reach the correct answer ⁶. A tool that helps enormously on some tasks and hurts on others produces exactly the kind of noisy, contradictory aggregate signal the labor data is currently showing, and it is a large part of why two careful research teams can look at the same underlying shift and describe it so differently. It is also why your own task list, done honestly, will usually tell you more than either headline.

That disagreement is not a reason to wait for certainty before acting on what you can already see. The task list from the previous section is worth revisiting every few months, since the ratio it describes is the part that actually moves.

See the benchmarks >