Estimated occupational exposure to AI varies substantially across models, shaping labour market forecasts, the identification of vulnerable sectors, and the design of reskilling initiatives
A worker who ready that her occupation is "highly exposed" to artificial intelligence may reasonably conclude that her job is in danger. A policymaker may use the same estimate to decide where to direct scarce training resources. But how reliable are these numbers? Recent evidence suggests: much less than many assume.
In a recent NBER working paper, me and my co-authors examine a widely used method for estimating occupational exposure to AI. We applied the same standard occupational task descriptions from the US Department of Labor to four leading frontier AI models—GPT-4, GPT-5, Gemini 2.5, and Claude 4.5—and asked each to assess exposure.
The results differed dramatically. Depending on the model used, the share of occupations classified as having “high direct exposure” to AI ranged from 14% to 51%. In other words, the estimated scale of AI risk more than tripled, even though the underlying occupational information was identical. The divergence does not stop there. When these exposure measures were linked to local employment data using standard empirical methods, the conclusions reversed. Under one model, higher AI exposure predicted employment losses. Under another, it predicted employment gains.
This is not simply a technical curiosity. These measures increasingly shape public debate and policy. International organisations, consulting firms, financial institutions, and governments rely on occupational AI exposure estimates to forecast labour market disruption, identify vulnerable sectors, and design reskilling initiatives.
If the same worker can be classified as “high-risk” or not depending solely on which AI model performs the assessment, confidence in precise headline numbers should be limited.
Why does this happen? A common reaction is to assume that better methods will eventually produce a single correct estimate. But the problem may be more fundamental.
AI models are not interchangeable measurement devices. They are trained on different data, designed with different objectives, and updated continuously. Their disagreement does not necessarily reflect temporary statistical noise. It may instead reveal genuine uncertainty about what “AI exposure” actually means.
Does exposure mean that some tasks can be partially automated? That most tasks can be replicated? That productivity will rise? That jobs will disappear? These are distinct questions, yet public discussions often collapse them into a single indicator.
This matters because labour market projections influence real decisions. Workers may reconsider training or career choices. Governments may channel public funds toward some sectors and away from others. A single estimate presented with unwarranted certainty risks misleading both.
None of this means that AI exposure measures should be abandoned. They remain useful indicators of where technological capabilities are advancing most rapidly. But they should be presented with greater caution.
Rather than reporting a single definitive number, researchers should show the range of plausible estimates across different measurement approaches. Policymakers should not base reskilling priorities, education planning, or labour market interventions on a single AI exposure estimate produced by one model. Public institutions should communicate uncertainty explicitly. When measurement is this unstable, acting as if the numbers are precise is not prudent policymaking, it is guesswork.
© Michelle Yin
Michelle Yin is Director and Associate Professor at Northwestern University and Founding Director of the Research and Innovation for Social and Economic Inclusion (RISEI) Lab
Please note:
We recognize that World of Labour articles may prompt discussion and possibly controversy. Opinion pieces, such as the one above, capture ideas and debates concisely, and anchor them with real-world examples. Opinions stated here do not necessarily reflect those of the LISER. A related piece was published on VoxEU.
Related World of Labour content:
https://wol.iza.org/articles/artificial-intelligence-and-labor-market-outcomes by Nick Drydakis
https://wol.iza.org/articles/how-is-new-technology-changing-job-design by Michael Gibbs Sergei Bazylik
https://wol.iza.org/articles/the-changing-nature-of-jobs-in-central-and-eastern-europe by Piotr Lewandowski
https://wol.iza.org/articles/who-owns-the-robots-rules-the-world by Richard B. Freeman
https://wol.iza.org/articles/the-gig-economy by Paul Oyer
Photo by skynesher on iStock