Opens in a new tab
needsahuman
Search ⌘KGet the data
Guide · updated 2 October 2026

What “AI exposure” really means

AI exposure is not the same as job loss. What researchers mean by exposure, how the main studies measure it, and how to read the numbers.

Key points

  • “Exposed to AI” means AI could affect some of a job’s tasks. It does not mean the job will disappear.
  • Different studies measure different things: what AI could do, what people actually use it for, or how well it does the work.
  • Exposure can mean AI replaces a task or helps a person do it. Most measured use so far is helping.
  • High-exposure jobs are often well paid and held by college graduates.
  • Always ask what a number measures before you worry about it.

Why the word causes confusion

You may have read that “a quarter of jobs are exposed to AI” and taken it to mean a quarter of jobs will go. That is not what the researchers mean.

In this research, exposure means that AI could change how some tasks in a job are done. The change might be AI doing the task, or AI helping a person do it faster. It says nothing on its own about whether the job shrinks, grows or simply changes.

The International Labour Organization put it plainly in its 2025 global index: “the figures reflect potential exposure, not actual job losses,” and “transformation, not replacement, is the most likely outcome.”

Three ways to measure exposure

1. What AI could do in theory

The best-known study is “GPTs are GPTs” by Eloundou and colleagues. They rated every task in the US government’s O*NET job database with a simple rubric:

  • No exposure (E0): AI does not meaningfully cut the time the task takes at the same quality.
  • Direct exposure (E1): a language model “can decrease the time required to complete the DWA or task by at least half (50%).”
  • LLM+ exposed (E2): the model alone would not halve the time, “but additional software could be developed on top of the LLM” that would.

Their headline: “around 80% of the U.S. workforce could have at least 10% of their work tasks affected,” and “approximately 19% of workers may see at least 50% of their tasks impacted.”

Notice what this does not say. The authors write that “technical feasibility does not guarantee labor productivity or automation outcomes,” and “we do not make predictions about the development or adoption timeline.”

An older approach, the AI Occupational Exposure measure by Felten, Raj and Seamans (2021), links AI progress to the abilities each job needs. A 2023 version for language models found that “top occupations exposed to language modeling include telemarketers and a variety of post-secondary teachers.”

2. What people actually use AI for

Newer studies look at real use instead of theory.

Microsoft Research analyzed “200k anonymized conversations with Microsoft Bing Copilot” from 2024 and scored each occupation by how much of its work AI was used for. Anthropic’s Economic Index does the same with conversations on its Claude assistant.

These usage measures are usually lower than theoretical ones, because people have not yet adopted AI for everything it could do. Anthropic’s March 2026 study found that “97% of the tasks observed” in its data “fall into categories rated as theoretically feasible.” In other words, real use is a subset of what theory predicts.

3. How well AI does the work

Exposure studies mostly ask can AI do this task? They rarely ask does it do it as well as a professional? That needs head-to-head tests, where experts compare AI’s work with people’s work without knowing which is which. Few jobs have that kind of test yet.

On NeedsAHuman, that is our second question, Is it better than a person?, and we grade the evidence from A to D because it is so uneven.

Exposure is not the same as replacement

The Microsoft researchers warn directly against reading their scores as a job-loss forecast:

“It is tempting to conclude that occupations that have high AI action applicability score will be automated and thus experience job or wage loss… This would be a mistake, as downstream consequences of new technologies are very hard to predict.”

They point to ATMs and bank tellers as an example of a technology that changed a job without simply removing it.

There is also a split between AI doing a task (automation) and AI helping a person (augmentation). In Anthropic’s first report, 57% of use was augmentation and 43% automation.

Who is most exposed

Exposure runs against old ideas about automation. Robots threatened factory jobs. AI language tools reach office work, and often better-paid office work.

Pew Research Center found that “in 2022, 19% of American workers were in jobs that are the most exposed to AI,” while “23% of workers have jobs that are the least exposed.” Workers in the most exposed jobs earned “$33 per hour, on average, compared with $20” for the least exposed. Pew adds: “We make no determination as to whether workers may lose their jobs as a result.”

Anthropic’s 2026 study found the most exposed workers “earn 47% more, on average.”

The ILO found that “25 per cent of global employment falls within occupations potentially exposed to GenAI, with higher shares in high-income countries (34 per cent).” It also found that the highest-risk jobs make up “9.6 per cent of female employment” in high-income countries, against “3.5 per cent” among men, because women are more often in clerical roles.

Does exposure predict anything?

Early signs say it may, a little. Anthropic compared its usage data with official job projections and found: “For every 10 percentage point increase in coverage, the BLS’s growth projection drops by 0.6 percentage points.” That is a small effect, and the same study found “no systematic increase in unemployment for highly exposed workers” so far.

Our guide AI and entry-level jobs covers the clearest signal yet: hiring of young workers in exposed jobs.

How to read any exposure number

When you see an exposure figure, ask five questions:

  1. Theory or use? “Could be affected” and “is being used for” are very different.
  2. Tasks or jobs? “10% of tasks exposed” is not “10% of jobs exposed”.
  3. Help or replace? Does it separate augmentation from automation?
  4. Quality? Does it say anything about how well AI does the work?
  5. When? Most exposure studies make no timeline claim at all.

How NeedsAHuman uses exposure

We combine all three kinds of evidence. Our Can AI do it? score blends observed use, a task-by-task capability rating and benchmarks. Our Is it better than a person? score adds quality, with an evidence grade. Only then do we estimate when a job could be replaced, always with a range.

Sources

How long will your job last?

Answer a few questions about how you actually do your job and get your own Still needs a human score, with the range, the risks and three things to do next.