Harvard Business ReviewNovember 2, 2026
Jay Fitzgerald, on research by Joseph Pacelli, Janet Gao, and Jun Oh
Why it matters. 2.9 million US workers analyzed, with resume language shifting to match the prevailing climate and that signaling influencing hiring outcomes even where predicted job fit was slightly lower. The study reads observed language against observed hiring, so the causal arrow is inferred rather than assigned. What it documents is a market that pays for phrasing. Candidates learned which words land this year and rewrote themselves accordingly, which is rational behavior and ruinous as a selection signal, because the skill it rewards is reading the room on a page nobody will ever check. The counter is not better resume screening. It is a step in the process where the candidate does the work and somebody who watched them do it goes on record.
ACL EMNLP 2026October 2026
Huy Nghiem, Phuong-Anh Nguyen-Le, Sy-Tuyen Ho, and Hal Daume III
Why it matters. Close to a million generated resume summaries, where the factual content stayed largely stable and the evaluative wording moved with race and gender name cues. Generated summaries in a research setup are not your ATS, and the effects sit in the tails rather than the average. The finding still lands on the exact thing most hiring audits are built to check. A fairness review that confirms the facts were carried over correctly will pass a summary that calls one candidate promising and another adequate on identical evidence. Framing is what a human reader acts on. If a model touches candidate text anywhere in your funnel, the audit has to read the adjectives, not just reconcile the dates.
arXiv preprintOctober 30, 2026
Nicolo Pagan, Michael Andreas Ebling, Stefania Ionescu, and Aniko Hannak
Why it matters. 5,500 AI-written cover letters and 414,000 model evaluations, with scores shifting according to which model drafted the application while the underlying CVs were held constant. Simulated pipeline, preprint, so hold the magnitudes loosely. The mechanism is the part worth acting on, and it echoes the self-preferencing result from two issues ago. A candidate's score is picking up their tool choice, which is a function of budget, habit and whatever their last employer paid for, and neither the applicant nor the recruiter can see it in the output. Everything written before the interview is now a measurement of the writing stack. Weight the part of your process the model cannot submit on the candidate's behalf.