ExplainerWorkNew
AI and jobs: how to read the numbers
Estimates of AI's effect on work range from task "exposure" to controlled productivity studies to company layoff statements, and they measure very different things.
Evidence: 2 sources · 2 research papers
The short version
- An "exposure" estimate says how many of a job's tasks AI could affect. It is not a forecast of job losses.
- Controlled studies have found productivity gains for some tasks, often larger for less experienced workers.
- When a company links layoffs to AI, that is the company's explanation, which may not be the only reason.
- National labour statistics change slowly and rarely isolate the effect of a single technology.
Few AI topics produce more confident numbers, and more confusion, than jobs. A headline that "40% of jobs" are affected and a headline that a study found a "14% productivity gain" are answering different questions.
Four kinds of evidence
1. Exposure estimates. Researchers break jobs into tasks and estimate how many an AI system could perform or assist with. The result is a measure of potential, not a prediction. A highly "exposed" job may change rather than disappear.
2. Controlled studies. These compare workers with and without AI on real or realistic tasks. Examples include a study of customer-support agents given an AI assistant and an experiment with writing tasks. Both found gains on the tasks measured. Such studies are strong evidence about those tasks, in those settings.
3. Surveys. Surveys of workers or employers show how people say they use AI and what they expect. They are useful for spotting trends, but answers depend on how questions are worded and who replies.
4. Company statements. When a company says AI let it cut jobs, or will, that is one explanation. Costs, demand, and strategy are often part of the story too.
Questions to ask
- Is this a measure of what AI could do, or what it did?
- Who was studied, and could the result apply to other jobs or countries?
- Was quality measured, or only speed?
- Who paid for the research?
How we cover work
We name the type of evidence behind each claim. We do not turn exposure estimates into job-loss predictions.
Limits and open questions
- Whether productivity gains in studies translate into higher wages, fewer jobs, or new kinds of work is still an open question.
Evidence and sources
- Preprint (not peer reviewed) Generative AI at Work
- Peer-reviewed study Experimental evidence on the productivity effects of generative artificial intelligence