Anthropic sees no clear AI-driven shock in unemployment

Anthropic, the company behind the AI assistant Claude, has compared what language models can theoretically do with how Claude is actually used in American professions. The result does not support a simple list of jobs that will soon disappear.

The study is based on tasks in the American database O*NET, previous estimates of language model capabilities and anonymized usage patterns from Claude. The researchers weigh automated use more heavily than when AI helps a human to perform the task.

The difference between possible and observed use is large. In computer and mathematics professions, language models were judged to theoretically be able to influence 94 percent of work tasks, while Claude was observed in 33 percent. For office administration, the corresponding theoretical proportion was 90 percent. It is a measure of tasks, not the percentage of jobs that disappear.

When the researchers compared more and less exposed occupations, they saw no systematic increase in unemployment since the end of 2022. They saw a possible decline in the rate at which 22-25-year-olds got jobs in more exposed occupations, but the result was close to the limit of statistical certainty and does not show that AI was the cause.

In a follow-up survey in June, nearly six in ten Claude users believed that AI could handle a greater portion of their work tasks within a year. The sample was not representative of the labor market. Computer and math occupations accounted for about 30 percent of the responses but about 4 percent of employment in the United States.

Both analyzes apply primarily to the United States and are based in part on the use of a single AI service. They therefore do not say exactly what is happening in Sweden or in professions where Claude is used less. Anthropic describes the work as an early way to follow changes, not as a sure forecast.

Därför spelar det roll

Charts of AI and professions are often circulated as lists of winners and losers. For Swedish employers and employees, a better question is which tasks actually change, who controls the results and whether new employees still get the opportunity to learn the job.

Det här kan du göra

1. Break down a role into concrete tasks before assessing how AI affects the job.
2. Measure actual usage, time savings, errors and need for human control instead of starting from a theoretical list.
3. In particular, track how junior employees receive training and entry-level roles as routine tasks are automated.