Analysis · 26 July 2026 · 7 min read

What the research actually agrees on about AI and jobs

We scored 961 occupations against 20 published research sources. Almost nothing clears the high risk threshold, and the estimates vary far more than the confident way they are usually quoted.

We built this site to answer one question for individuals: how exposed is my job. Doing that meant reading 20 published research sources on AI and work, extracting their occupation level estimates, and reconciling them into one number per role.

Reconciling them turned out to be the interesting part. Here is what the research base actually looks like once you line it up side by side, including the finding that surprised us most.

Every figure and table below is generated from the live index rather than written by hand, so this post cannot drift away from the ranking itself as new research lands.

Almost nothing scores as high risk

Our bands are simple. Below 40 is lower risk, 40 to 69 is transitioning, 70 and above is high risk.

Across 961 occupations backed by at least two sources each, exactly zero clears 70. The mean score is 31. The median is 29.

That is not the picture you get from headlines, and it is not the picture we expected when we started ingesting papers. The single highest scoring occupation in the whole index, Customer Service Representatives, sits at 68.

Read carefully, the aggregate research consensus is not saying that any specific occupation is close to fully automatable today. It is saying that a meaningful slice of tasks in a minority of occupations is now within reach of the tools, and that the slice is largest in office and text heavy work.

Three caveats keep this honest. Weighting is ours: we weight more recent and higher confidence estimates more heavily, and a different scheme would move individual numbers. These are measures of task overlap, not forecasts of employment, so a score of 30 does not mean a job is 70 percent safe forever. And the estimates are extracted from each paper automatically rather than hand coded by a researcher, which adds error we can bound but not eliminate. The full set of limitations is on our data page.

What the sources broadly agree on

Three things show up consistently once you compare estimates for the same occupation across different studies.

Direction. Sources disagree about magnitude far more than direction. Where one study calls an occupation more exposed than average, others usually agree it is above average, even when their numbers differ a lot. Administrative, clerical and text handling roles are consistently placed above the middle. Physical, hands on and unpredictable environment roles are consistently placed below it.

Task content beats industry. Every source that works at task level lands in a similar place: what matters is the mix of activities inside a role, not the sector it sits in. Our own data shows this starkly, with healthcare occupations appearing both near the top and near the bottom of the ranking.

Exposure is not displacement. Nearly every serious source in this base is careful to say it measures overlap between job tasks and model capability, not job losses. The gap between those two ideas is where most misleading coverage lives.

Where the estimates diverge, and why

This is the part we think deserves more attention, with one important caveat we want to put before the numbers rather than after them.

For every occupation with more than one source, we can measure the spread: the gap between the lowest and highest individual estimate before aggregation. Across our dataset, the median spread is 34 points on a 100 point scale. The mean is 33. And 567 of 961 occupations have a spread of 30 points or more.

That spread is not purely disagreement. Part of it is genuine: studies looking at the same work reach different conclusions. But part of it is that the sources are not all measuring the same quantity. A task overlap score, a wage weighted exposure estimate and a survey of what employers say they plan to do are three different constructs, and our pipeline places them on one 0 to 100 scale so they can be combined. That comparison is a modelling choice, not something the underlying research endorses, and it inflates the apparent disagreement by an amount we cannot cleanly separate out.

We are stating it plainly because it cuts against the most quotable number on this page. Treat the spread as an upper bound on disagreement rather than a measurement of it.

The widest divergences among occupations with at least four sources behind them:

Look at the range on any row here. The same occupation, described by published and broadly contemporary research, lands at opposite ends of the scale depending on which study you read. Notice too how many of those rows are postsecondary teaching roles carrying identical ranges, which is its own tell: at least one source is stamping one value across a whole family of job titles rather than assessing them separately.

If you have ever read two credible articles about AI and your own field that seemed to describe different universes, this is a large part of why. But the honest reading is not that one of them is simply wrong. Several things are mixed together in a gap that size:

• The sources define the occupation differently, and a broad title covers very different day to day work.
• They measure different quantities, as above, which our scale flattens together.
• They were written at different points in a fast moving field. A 2023 estimate and a 2026 estimate are not describing the same technology.
• They cover different countries and labour markets.
• Our own extraction adds error on top, since the estimates are pulled out of each paper automatically rather than hand coded.

So the takeaway is not "the research is a mess". It is that a single number lifted from a single study, which is how nearly all coverage of this topic works, carries far more uncertainty than the confident way it is usually quoted.

What that means for reading any single number

Three practical conclusions, including one that is inconvenient for us.

Be suspicious of any single confident figure, including a single study that says your job is 85 percent automatable, and including a single number on a website. The honest version of this research is a range with disagreement inside it, which is exactly why every occupation page here lists the individual sources behind its score.

A consensus is more useful than a favourite study, but it is still a summary. Aggregating reduces the influence of any one outlier. It does not make the underlying uncertainty disappear, and averaging can flatten a genuine signal from the newest study.

The direction is more trustworthy than the magnitude. If your role is text and screen heavy, the sources agree you are more exposed than average, even though they will argue about whether that means 40 or 70. That directional signal is actionable. The precise number is not.

The sources

The 20 publications behind these scores range from peer reviewed academic work to government statistics and consultancy analysis. By number of occupations covered, the largest contributors are:

Coverage is uneven, and that is the limitation we would flag hardest. The single widest source, AI Occupational Exposure to Language Modeling, contributes an estimate to 940 of the 961 published occupations. At the other end, 13 occupations (1 percent) rest on exactly two sources, while the best corroborated carry 14.

An occupation backed by two sources is not held to the same standard as one backed by ten, which is why the source count sits on every occupation page and in the divergence table above.

Occupations scored by only one source are excluded from the public index entirely. That is why the index covers 961 occupations rather than every role in our database.

You can see the full ranking, with the source list for each occupation, in the AI Job Risk Index, and read how the aggregation works in our methodology.

Figures come from the AI Job Risk Index and were current when this was published. Scores change as new research is ingested, so the index is always the live version. See how scoring works. Informational guidance based on published research, not professional career or financial advice.