The Census Bureau measured a 13% earnings drop for AI-exposed graduates

Working paper CES-26-56 by Cody Orr, Lee C. Tucker and Lawrence Warren uses administrative records, not a survey and not a vendor benchmark, to track college graduates since ChatGPT shipped in late 2022 (US Census Bureau). The most AI-exposed decile of majors saw initial employment likelihood fall five percentage points. Full-quarter initial earnings fell 13 percent. The authors describe the magnitude as comparable to graduating into a large recession.

The mechanism finding is the one that changed my thinking. About half the earnings decline comes from graduates shifting into lower-wage sectors, restaurants and retail, rather than from lower pay inside the same field. The popular version of this argument assumes AI exposure shows up as wage pressure on the job you trained for. The data says it shows up as not getting that job at all.

I've been skeptical of displacement claims for two years because almost every one of them ran on anecdote, vendor marketing, or a survey where people guessed about their own future. This is administrative earnings data from the agency that collects it. It doesn't prove causation, and the authors are careful about that, but a five-point employment gap concentrated in the most exposed decile is not the kind of pattern that shows up from noise.

Two other things today sit next to it. A hiring manager posted an Ask HN thread saying roughly 80% of developer candidates now tell him they aren't writing much code themselves, they're directing agents, and asking whether leetcode and system design interviews measure anything anymore (Hacker News). The top replies converged on a two-stage format: a simple take-home, then a pairing session where the candidate extends their own submission without AI. Extending your own code is where it becomes obvious whether you understood it.

And AWS moved Amazon Connect Talent to general availability on September 17, which puts AI agents in front of candidates running interviews and assessments on demand, returning scored summaries with transcripts and per-competency evidence (AWS). Built on Amazon's internal hiring systems, same playbook that turned in-house contact center tech into Amazon Connect.

So the entry-level cohort is getting squeezed on the demand side by the tooling and screened on the supply side by an agent. I don't have a clean action item for this one. If you're hiring juniors, the pairing-on-your-own-code format is the most useful concrete idea I saw today, because it tests the thing that the Census data implies is becoming scarce: understanding what you produced.