By Justin Lahart
The numbers that researchers have been using to predict which jobs could be lost to artificial intelligence may be unreliable. One of the culprits could be AI itself.
Worries about how AI will affect the job market have become one of the most pressing economic questions of our time, and understanding the limits of any forecasts is crucial. Policymakers want to know which workers could be displaced and need support. Schools, students and parents want to know which careers might be "AI-proof."
To gauge which jobs might be at risk, economists have built "exposure scores" using a task-based framework. The Labor Department maintains a database of what workers in different occupations actually do: Bakers need to mix dough and put it in the oven, financial analysts need to assess companies, and so on. Researchers try to determine which of those tasks AI could significantly speed up.
The larger the share of tasks that AI could do, the more exposed that job is. These scores have proliferated widely, appearing in research notes, consultancy white papers and policy-advocacy reports.
Researchers have three main ways to build exposure scores: human raters who manually assess how well AI handles different tasks; surveys of workers who use AI platforms; or AI models themselves. Each has drawbacks. Human ratings can be highly subjective. Worker surveys reflect only users of one platform who may not represent the broader workforce.
The third option -- having AI rank which tasks are most exposed to AI -- has its own, unique set of problems.
That is one of the takeaways from a new study posted last month on the National Bureau of Economic Research website. Economists Michelle Yin and Hoa Vu of Northwestern University and Claudia Persico of American University asked three AI models -- OpenAI's ChatGPT-5, Google Deepmind's Gemini 2.5 and Anthropic's Claude 4.5 -- which jobs were most exposed to AI, and often got widely varying answers.
The researchers found that Claude rated accountants as highly vulnerable to AI, for example, while Gemini assigned them a much lower exposure ranking. Other occupations where the models disagreed about vulnerability to AI included advertising managers and chief executives. ChatGPT and Gemini were the most in sync, but they still disagreed about a quarter of the time.
Some of these differences came down to differences in the models, but the economists also found evidence that the readings were partly shaped by which workers were already using AI. Early adopters such as financial analysts use AI heavily, generating more training data for future AI models, and that in turn feeds into how the models rank that profession.
The problem, the economists say, is that some policymakers and employers might not be taking these scores with an appropriate grain of salt. (Their study, as a working paper, hasn't yet been peer reviewed.)
To be fair, disagreement among different versions of a rapidly emerging technology isn't necessarily surprising. What's more, it isn't clear that AI is any better or worse at measuring exposure than the other common methods.
As a first step, the economists think researchers should look across a variety of models instead of just one, and be forthright about how uncertain AI-generated exposure readings might be. Ultimately, the researchers say, surveys of how AI is actually being implemented across the economy, and which tasks it is being used for, might yield better answers on how exposed different jobs are.
"I personally would not rely on just one measure to say, 'Oh, I should change my job,' or 'I should change my kid's major,'" said Yin, one of the study's authors.
Write to Justin Lahart at Justin.Lahart@wsj.com
(END) Dow Jones Newswires
May 10, 2026 12:00 ET (16:00 GMT)
Copyright (c) 2026 Dow Jones & Company, Inc.
Comments