How Accurate Are AI Job Loss Predictions? Three Checks in a Minute

How accurate are AI job loss predictions? Most of the specific ones you see in a headline are not measurements of anything happening in the economy. They are one model's guess about a job title, dressed up as fact. A few of the recurring claims trace back to real, reachable research. Most do not survive the trip from a careful paper to a confident headline. Below are three checks you can run on any AI job risk claim in under a minute, and what each one is built to catch.

You have probably seen the genre: a list claiming dozens of jobs face near total automation, shared everywhere without a single link to where the number came from. This is not an argument that every such claim is invented. It is a way to tell which ones are worth believing.

Check one: exposure, or displacement

The first check is whether the claim is even measuring what you think it is measuring. Exposure means a task could technically be handled by AI under current conditions. Displacement means an actual person loses an actual job. The International Labour Organization is blunt about the gap between the two: exposure indicators reveal technological susceptibility, not labour market outcomes, and cannot be interpreted as predictions of job displacement. A headline that slides from one word to the other, usually inside a single sentence, has stopped describing the research it is citing.

Check two: evidence, or a language model's guess

The second check is where the number actually came from. Many exposure scores are built by asking a language model how automatable a task sounds, with no supporting evidence attached and no way to inspect the reasoning behind the answer. A 2026 paper tested an alternative on 18,796 occupation task pairs drawn from the O*NET database: it gave models real news articles and academic abstracts as evidence, then had human reviewers compare the evidence grounded verdicts against the ungrounded guesses. The evidence grounded version won in 72 percent of the cases where the two disagreed. A claim that cannot point to any evidence behind its number is the older, weaker kind of guess, dressed up as data.

Check three: does the number predict anything real

The third check matters most, because it asks whether the score correlates with something that actually happened. A 2025 study in PNAS Nexus matched a decade of monthly unemployment claims, 2010 through 2020, against several published AI exposure scores, occupation by occupation, state by state. The single best exposure score explained only 10.7 percent of the real variation in unemployment risk, the same study found. Combining many different scores together did better, adding 18.1 percentage points on top of a baseline that already accounted for skills, education, and region. A single published score, taken alone, is a weak predictor of whether real people end up filing for unemployment.

What to do with a headline that fails these checks

Most headlines fail at least one of the three checks, and the more dramatic the figure, the more likely it slid from exposure toward displacement somewhere between the paper and the retelling. None of this means the underlying research is worthless. It means a job title score was never built to describe your own week. Our checklist for reading your own week instead of a job title exists for that reason, and it sits next to how a job level score actually gets built in the first place. Both point at the same underlying gap: the judgement, communication, and decision making skills that resist automation are exactly the part a task list was never built to record. If you want a read produced from your own answers instead of a stranger's guess about your job title, the check scores six domains directly and takes about seven minutes.