While there have been a lot of varied claims about the impact of AI on the workforce, no one can currently say with certainty what kind of jobs will be most affected by the new technology. In order to put the debate to an end, researchers asked some of the world’s most advanced AI models which jobs are most vulnerable to artificial intelligence.
However, the study instead found that these AI-generated predictions, often referred to as “exposure scores”, may be highly unreliable.
According to a new working paper reported by The Wall Street Journal, economists Michelle Yin and Hoa Vu of Northwestern University, along with Claudia Persico of American University, tested OpenAI’s ChatGPT-5, Google DeepMind’s Gemini 2.5, and Anthropic’s Claude 4.5 to see how each model ranked professions based on their exposure to AI.
AI models couldn’t agree on which jobs are most at risk:
The working paper, which was published on the National Bureau of Economic Research website, found that the AI models frequently disagreed on which jobs were most at risk.
The study found that these AI models disagreed the most on supervisory roles and jobs which combine cognitive and physical tasks. Meanwhile, the models seemed to be more aligned when ranking physical jobs.
The researchers found that Claude rated accountants as highly vulnerable to AI automation, while Gemini assigned the profession a much lower exposure ranking. The models also disagreed on the vulnerability of advertising managers and chief executives.
They also found that the estimates of the impact of AI on employment could change drastically based on the AI model used.
For instance, they found that ChatGPT-5 or Gemini 2.5 would conclude there is no statistically significant association between AI exposure and employment. However, a researcher running the exact same data through Claude 4.5 would report a statistically significant negative relationship.
“The same data, the same specification, and the same time period produce event study trajectories that differ in magnitude by a factor of two, solely because a different model assigned the exposure scores,” the researchers wrote.
The researchers also found evidence that AI adoption itself may shape future exposure scores. They say that occupations already using AI heavily, such as financial analysis and digital office work, generate more training data for newer models, potentially influencing how future systems rate those professions.
Do not rely on a single AI model:
The study warned researchers against treating any single AI-generated exposure score as definitive, especially for high-stakes policy decisions around education, hiring, and workforce planning.
Michelle Yin, an economist at Northwestern University and one of the study’s authors, told WSJ, “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,’”
Famously, Anthropic’s CEO Dario Amodei has predicted that AI will lead to the loss of around 50% of white-collar jobs in the next 2-5 years.
Meanwhile, Nvidia CEO Jensen Huang has cautioned against such claims and instead noted that human jobs would be taken over by a human using AI.
Source link
#Researchers #asked #ChatGPT #Gemini #Claude #jobs #exposed #chatbots #wildly #diagree #Mint
