An AI job-risk score changed with the platform supplying its data

What the researchers held fixed

Yin and Ogut used a 2015–2024 American Community Survey panel of about 13.1 million person-years. They held the employment outcome, comparison design, task-capability rubric and estimator fixed. They then rebuilt an occupation exposure score with conversation shares from different Claude consumer and enterprise releases and Microsoft Copilot. The exercise tests whether platform logs act like a stable workforce exposure measure.

The result changed with the users observed

Across tested platform inputs, the post-ChatGPT employment coefficient varied by a factor of 1.9. Within one vendor, consumer and enterprise channels gave opposite signs at each observed release. After reweighting platform occupation shares to U.S. workforce shares, the tested coefficients shrank by 42–93%. That percentage describes coefficient attenuation in this diagnostic; it is not a change in employment.

What this does and does not change

People using an AI platform are not a random sample of workers. Conversation logs reveal useful task activity among users, but they may miss nonusers and workers whose tasks have already changed. Reweighting removes one visible difference in occupation mix; it does not remove within-occupation selection or other causes of employment change. The authors do not identify a causal AI effect on jobs. Their result applies to scores built with platform conversation shares, not every exposure index or direct employer-adoption measure.

Our earlier entry-level hiring warnings and broad U.S. tracker use different populations and measures. This paper does not cancel either one. It raises a sharper question for any platform-based claim: would the result survive another platform, a workforce weighting or observed employer use?

Before using an exposure headline

1. Was the score built from what AI could do, what workers report doing, or conversations on one platform?
2. If platform logs were used, whose work is absent from that platform?
3. Did the study measure jobs or wages, and did its result survive another exposure measure?

Your employer's actual changes to tasks, training, hiring and staffing still matter more to your own next question than a platform rank alone. The survey and the rank cannot predict your personal outcome.