[Digital Today reporter Hyunwoo Choo (추현우)] Anthropic on March 5 (local time) released a report analysing AI’s impact on the labour market, titled “AI’s Impact on the Labor Market: New Evidence and a Novel Measurement Approach.”

The report’s core is a new AI replacement-risk indicator called “Observed Exposure.” While earlier studies relied only on AI’s theoretical capabilities, the index also reflects actual usage data. It assigns greater weight to automation and work-related use, designed to capture real job impacts more precisely.

The researchers calculated the index by combining data on about 800 occupations from the U.S. occupational information database O*NET, usage data from the Anthropic Economic Index, and task-by-task exposure estimates from the Eloundou research team.

A notable point in the report is that a large gap remains between tasks AI can perform in theory and actual use.

For example, in computer and mathematics fields, AI can theoretically handle 94 percent of total work, but based on actual Claude usage it accounted for only 33 percent. Office and administrative roles also had a theoretical exposure of 90 percent, but actual use was far lower. The researchers analysed legal restrictions, software requirements and human verification processes as factors delaying adoption.

By occupation, observed exposure was highest for computer programmers at 74.5 percent. They were followed by customer service representatives at 70.1 percent, data entry keyers at 67.1 percent, medical records specialists at 66.7 percent and market research analysts at 64.8 percent. Financial and investment analysts at 57.2 percent, software quality assurance analysts at 51.9 percent and information security analysts at 48.6 percent also ranked high.

By contrast, frontline jobs such as cooks, bartenders, motorcycle mechanics and lifeguards were classified as having 0 percent exposure. That means AI is not used at all. About 30 percent of all workers fall into this category.

Compared with U.S. Bureau of Labor Statistics (BLS) employment projections, the report found a correlation in which every 10 percentage-point increase in AI exposure was associated with a 0.6 percentage-point lower projected employment growth rate for 2024 to 2034. Characteristics of workers in highly exposed occupations also drew attention. As of August to October 2022, just before ChatGPT’s release, workers in the top 25 percent of exposure had average hourly wages 47 percent higher than those in non-exposed occupations. The share of women was 16 percentage points higher, the share of white workers was 11 percentage points higher, and the proportion with graduate-level education or above was about four times higher, at 17.4 percent versus 4.5 percent.

However, from ChatGPT’s launch in November 2022 to the present, increases in unemployment among highly exposed occupations were not statistically significant. The researchers said that comparing unemployment-rate trends for the top 25 percent exposure group and the non-exposed group showed almost no change in the gap between the two.

This suggests AI’s impact could emerge gradually, like the spread of the internet or China-driven trade shocks, rather than abruptly like COVID-19.

Still, an exceptional signal was detected in one area. New hiring in AI-exposed occupations among 22 to 25-year-olds fell 14 percent from 2022. While the monthly employment rate for non-exposed occupations was stable at 2 percent, the employment rate for highly exposed occupations was about 0.5 percentage points lower. No such decline was seen among those older than 25. The researchers said this could indicate an initial employment impact from AI, but alternative interpretations are also possible, including returning to study instead of taking a job or moving to other occupations.

The researchers said they plan to update the analysis continuously as new employment and AI usage data accumulate. The report is available on the Anthropic Research website (anthropic.com/research).