Human economists say the use of AI is far from reaching its full potential to disrupt labor markets.
Using the new scale, we found that the five most at-risk occupations are computer programmers, customer service representatives, data entry keystrokers, medical records specialists, and market research analysts and marketing specialists.
Economists Maxim Masenkov and Peter McCrory write that AI has not yet had a significant impact on unemployment rates for workers in these at-risk occupations. They said there is “suggestive evidence” that employment of young workers in these sectors is slowing.
Our measure, Observed Exposure, compares the tasks that LLMs could theoretically perform with the tasks that people actually use Claude in the workplace.
It turns out that practical usage is far from its theoretical capabilities. pic.twitter.com/cnbkRNMu26
— Peter McCrory (@PeterMcCrory) March 6, 2026
Masenkov and McCrory also wrote that there are many tasks, and in some cases entire jobs, that AI cannot perform, such as arguing legal arguments in court.
“Of course, many tasks remain beyond the reach of AI, from physical agricultural tasks such as pruning trees and operating farm machinery to legal tasks such as representing clients in court,” the pair wrote.
The core of Massenkoff and McCrory’s paper proposes a new way to measure AI replacement risk that combines real-world data about Claude usage with other factors, such as the tasks that AI could theoretically do.
Anthropic publishes actual data on Claude usage in each state and Washington DC through the Anthropic Economy Index.
Doing so will make it easier to more reliably identify economic disruption in real time and “identify the most vulnerable jobs before displacement becomes apparent,” they said.
“While this approach cannot capture all the channels through which AI may reshape the labor market, by laying this groundwork now, before any meaningful impact occurs, we hope that future findings will identify economic disruptions more reliably than post-mortem analysis,” they write.
This measure, which they call “Observed Exposure,” shows how far LLMs have to go to disrupt specific jobs that AI could theoretically replace or augment.
“For example, Claude currently covers only 33% of all tasks in the Computers and Mathematics category,” they write.
Dario Amodei warns about future of white-collar jobs
Anthropic CEO Dario Amodei has repeatedly warned of job losses due to AI. He said up to half of entry-level white-collar jobs could be replaced by AI in the next one to five years. Amodei remains steadfast in his views, even as industry insiders, including OpenAI CEO Sam Altman, question his outlook.
Massenkoff and McCrory’s findings are consistent with a growing consensus that AI has the potential to eliminate most entry-level software engineering jobs. One of Anthropic’s biggest uses for Claude is coding.
Claude Code creator Boris Cherny recently said he expects the software engineer title to “start to disappear” in 2026.
xAI CEO Elon Musk said last year that “things that physically move atoms” will outlast AI disruption. Humaneconomists found that the least exposed occupations include cooks, motorcycle mechanics, lifeguards, bartenders, and dishwashers.
It’s worth noting that blanket predictions about AI’s job destruction aren’t necessarily old-fashioned.
Jeffrey Hinton, the so-called “godfather of AI,” said in 2016 that “people should stop training radiologists now” and that AI will surpass humans in the field within five years. Ten years later, radiologists remain in demand. Hinton told the New York Times in 2025 that while he was right about the direction of AI’s progress, his predictions were too broad and the timing was off.
And the disruption of AI will not have the same impact on everyone, the human economists write.
Based on U.S. Census Bureau data from the three months before ChatGPT was released, the economists found that “workers in the most at-risk occupations are more likely to be older, female, more highly educated, and higher-paid.”
