What Is AI Exposure?
AI exposure is a measure of potential task impact. It is not a jobs-lost number.
The short answer
AI exposure means that a particular AI capability could affect some of the tasks connected to a job or occupation. It is usually a modelled estimate based on task descriptions and an assessment of what the technology might do.
That is useful information. It is also much narrower than the way the number is often presented.
Exposure does not, by itself, tell us that workers are using the system, that employers have adopted it, that productivity has improved, or that jobs have disappeared. It does not provide an adoption timetable or a wage forecast.
The distinction is the whole point. Computers have enough trouble keeping categories separate without humans merging them in a headline.
Exposure starts with tasks
A job is made up of many tasks. An accountant may prepare records, check unusual entries, explain results, answer questions, and take responsibility for a decision. A software developer may write code, understand an existing system, test changes, investigate failures, and communicate with other people.
An exposure study asks which of those tasks might be affected by a capability. It may then aggregate the task estimates to produce an occupation-level figure.
That does not mean the occupation will be automated. A task can be affected while the worker remains responsible for directing it, checking it, correcting it, or using the result in a larger process.
Exposure can lead to augmentation, transformation, adoption, or displacement. The exposure estimate alone cannot tell us which outcome will occur.
What the widely quoted numbers mean
Eloundou, Manning, Mishkin, and Rock’s 2023 paper, GPTs are GPTs, estimated that roughly 80% of the US workforce could have at least 10% of its tasks affected by large language models, while roughly 19% could have at least half of its tasks affected.
Those are technical-exposure estimates. They describe potential task impact under the paper’s definitions and method. They are not observations that 80% of workers are using AI, and they do not mean 19% of workers will lose their jobs.
The paper is explicit about the gap between technical feasibility and what happens in the real world. Adoption depends on cost, reliability, workflow integration, organisational choices, and other constraints. Labour outcomes require different evidence.
The figures are also bounded by their setting: a US-focused analysis, task definitions, source data, model assumptions, and the version of the technology examined. A percentage from one method should not be converted mechanically into a universal forecast.
Keep the evidence categories separate
When reading an AI claim, separate these questions:
- Capability: Could the system perform a related task?
- Exposure: How much potential task impact does a model estimate?
- Use: Are people actually using the system for that task?
- Adoption: Have organisations built it into regular work?
- Productivity: Did speed, cost, or quality change in a measured setting?
- Labour outcome: Did employment, hiring, hours, wages, or earnings change?
- Causation: Does the research show that AI caused the change?
A study can answer one of these questions without answering the others. A capability demonstration is not an adoption survey. An exposure index is not an employment study. A productivity result on one task is not an economy-wide labour forecast. The evidence chain is useful precisely because it prevents a plausible first step from being mistaken for the final outcome.
The project’s broader evidence baseline records both employment pressure in some exposed work and counterevidence showing that adoption can occur without large short-term changes in earnings or hours. That is not a contradiction. It is what happens when different studies measure different parts of the chain.
What AI exposure does not prove
An exposure percentage does not prove:
- that an entire occupation will disappear;
- that employers will adopt the capability;
- that workers will be replaced rather than supported or reassigned;
- that productivity will rise in every workplace;
- that wages, hours, hiring, or employment will change;
- that any observed labour-market change was caused by AI.
It also does not tell us who captures any productivity gain. That may depend on prices, staffing decisions, bargaining power, business strategy, and public policy. Those are separate questions requiring separate evidence.
A practical test for exposure claims
Before accepting a percentage, ask:
- What capability and which tasks were assessed?
- Who was included, where, and when?
- Is the number about technical potential, actual use, adoption, productivity, or an observed labour outcome?
If the claim does not answer those questions, treat the number as a starting point rather than a conclusion.
For a related example of why per-unit efficiency and total use must be measured separately, read What Is Jevons’ Paradox? and More Efficient AI Does Not Guarantee Lower Total Use.
Bottom line
AI exposure is best understood as estimated potential for an AI capability to affect identified work tasks. It can help us identify where to look more closely. It cannot, on its own, tell us what workers, employers, or the labour market will do next.
Research cutoff: 2026-08-03. Last reviewed: 2026-08-03.