Predictions about AI and jobs range from mass unemployment to a productivity boom. The research so far supports something less dramatic and more specific: AI is changing tasks within jobs faster than it is eliminating whole occupations, and the early effects are concentrated among less experienced workers in exposed fields.
Exposure is not the same as replacement
A widely cited 2023 study by researchers at OpenAI and the University of Pennsylvania estimated that about 80% of U.S. workers have at least 10% of their tasks “exposed” to large language models, and about 19% have at least half their tasks exposed. Exposure means a model could speed up a task, not that a job disappears. Many exposed jobs, such as writers, analysts, paralegals and programmers, involve tasks AI helps with and tasks it does not.
What productivity studies found
- Customer support. A study of more than 5,000 support agents at a software company found that an AI assistant raised the number of issues resolved per hour by about 14% on average, and by about a third for the least experienced agents. The most experienced agents saw little gain.
- Writing and consulting tasks. Controlled experiments with professionals, including a study with Boston Consulting Group staff, found faster and higher-rated work on tasks within the model’s abilities, and worse results when people relied on it for tasks outside them.
- Software. Experiments with coding assistants show large speedups on well-defined tasks, while studies of experienced developers on their own complex codebases have found much smaller gains, and sometimes none.
A consistent pattern: AI narrows the gap between novices and experts on tasks it handles well.
What the employment data show
Through 2025, analyses of the overall U.S. labor market, including one from the Yale Budget Lab, found no broad disruption in the mix of occupations since ChatGPT’s release beyond the pace of earlier technology shifts. Research from Stanford’s Digital Economy Lab using payroll data found a different signal underneath: employment for workers aged 22 to 25 in the most AI-exposed occupations, such as software development and customer service, fell relative to less exposed jobs, while older workers in the same occupations were not affected the same way. Entry-level hiring, not mass layoffs, is where economists are watching most closely.
History is a guide, with limits
Economists including MIT’s David Autor have found that most of today’s employment is in job types that did not exist in 1940. Technology has repeatedly destroyed some tasks and created new ones. The open question is whether new work will arrive as fast, and pay as well, as the work AI absorbs, and whether the gains will be shared. Those are policy choices as much as technical ones.
How to read AI jobs headlines
- Ask whether a number measures exposure, adoption, productivity or actual job losses. They are often confused.
- Check whether a company’s layoff announcement attributes cuts to AI with evidence, or just mentions it.
- Prefer studies with comparison groups over surveys of what executives expect to do.