AI job-risk scores from chat logs can confuse platform popularity with real workforce exposure.

AI labor studies may be measuring platform adoption more than job exposure.

i.e. AI exposure scores from chat logs mostly show who uses each platform, not just whose work AI can change.

The main finding is that platform-based measures often overrepresent computer, and office jobs while underrepresenting food, transport, production, and manual service jobs.

When the authors reweight the data to match real workforce job shares, the estimated employment effects shrink by 42% to 93%, and some results become close to zero.

----

Link – arxiv. org/abs/2605.21743

Title: "Who Uses AI? Platform Selection and the Measurement of Occupational AI Exposure"