A second Census study raises the stakes for AI-exposed graduates. It still is not a final verdict.

What was measured

The study links postsecondary and administrative employment records for roughly 6.7 million bachelor’s graduates from a subset of U.S. institutions. It assigns each major a pre-ChatGPT task-exposure score from the occupations its graduates typically entered, then compares outcome changes for the top exposure decile with the bottom six deciles around late 2022.

The result

In the authors’ regression-adjusted estimates, the most-exposed majors were five percentage points less likely to have initial employment and had 13% lower full-quarter initial earnings. About half of the earnings gap came from lower earnings within industries; the rest came from movement away from higher-paying sectors toward lower-paying sectors such as retail and accommodation and food services. Effects became smaller later in the early career, but did not fully disappear in the available window.

Why the AI attribution remains incomplete

Exposure is not observed adoption. The study's causal reading requires that, without ChatGPT, highly exposed majors would have moved like the less-exposed comparison group after 2022. The authors test work-from-home, graduate supply and other explanations, but state that their design cannot rule out every time-varying shock to particular fields. Its participating institutions also do not represent every U.S. college or graduate.

A broad-market cross-check points elsewhere

A Federal Reserve analysis of U.S. Lightcast job postings finds little indication of a distinct AI-driven decline in broad labor demand since ChatGPT, despite slower overall hiring. That does not erase the graduate result: postings are not hires, and the analysis is not designed to isolate new graduates or majors. It does block a lazy leap from a concentrated entry signal to a proven whole-market collapse.

Why exposure cannot stand in for use

New task-level U.S. survey data find that AI use reaches more than 80% of occupations, but is shallow in most. The same occupation can contain regular users and non-users. That is why an exposure score, an adoption measure and an employment outcome must stay separate.

What would change this conclusion

• Later cohorts showing whether graduates catch up, fall further behind or move into different work.
• Studies that combine observed employer adoption with entry hiring, wages and stronger controls for field-specific shocks.
• Evidence explaining why broad posting patterns and concentrated early-career outcomes diverge.