Comment on “AI Self-preferencing in Algorithmic Hiring: Empirical Evidence and Insights”

End-to-end test of /critical-ai-publish: fresh OA empirical paper sourced by the pipeline; full-text critique span-grounded to the ar5iv source store.

(see field)

A controlled resume-correspondence experiment (2,245 human-written resumes, 24 occupations, 9 LLM evaluators, 18 Prolific annotators for ground truth, conditional logistic regression + a simulated hiring pipeline) testing whether LLM screeners favor resumes generated by themselves. Headline: 68-88% self-preference bias and a 23-60% shortlisting advantage for same-LLM candidates.

Construct validity of the key 'human-written' baseline is questionable. The summaries are scraped from LiveCareer.com, a commercial resume-BUILDING platform that supplies templates, examples, and writing assistance; the paper asserts they are 'human-written' solely on the basis of pre-LLM timing, without verifying they are unassisted natural prose. If many LiveCareer summaries are template- or expert-derived (often deliberately generic or keyword-stuffed), the 'human vs AI' contrast partly measures naturalistic-human vs polished-AI style, which can bias the self-preference estimate and weaken the 'against human-written resumes is particularly substantial' claim. This is a real construct-validity gap, though its directional effect on the estimate is not established.

The most striking headline finding, a '100% equal opportunity self-preference bias,' rests on only 30 human-annotated resume pairs, with ground-truth quality labels derived from three annotators per condition aggregated via 10,000 bootstrap resamples. Bootstrapping cannot create information beyond the 30 underlying observations or 3 raters; reporting it as a clean 100% on such thin annotation overstates precision and stability. A clean 100% with n=30 is statistically fragile, and although the authors disclose the sample-size caveat, the result is still foregrounded as a primary contribution.

The labor-market impact (23-60% more likely to be shortlisted) is generated entirely by a forced-choice / simulated screening pipeline (five human vs five evaluator-generated resumes competing for four slots, forced ranked output), not observed hiring behavior, yet is framed as real-world 'labor market impact.' Real employers typically score single resumes against a bar rather than make head-to-head A/B picks within a stacked pool, so the forced binary/ranked choice likely amplifies any preference relative to field screening. The figure should be read as an upper bound from a forced-choice design rather than a field estimate of shortlisting effects.

The core phenomenon is genuinely robust and the experiment is, for a CS-adjacent social-science preprint, unusually disciplined. The effect appears consistently across nine models spanning closed and open source, with large and statistically significant conditional-logistic coefficients (e.g., GPT-4o 2.709***, 2,245 pairs / 4,490 observations), and the design controls several obvious artifacts a weaker paper would miss: ordering is counterbalanced, verbosity is constrained to the human-summary interquartile length range, and all non-summary content is held identical within a pair so the manipulation is tightly localized. The authors use blinded human annotators to establish quality ground truth, are candid about the binding limitation (explicitly flagging the 30-pair annotation), and propose low-cost mitigations that cut bias by more than half. Even granting the self-recognition-versus-style ambiguity, the practical upshot is similar: whatever the precise mechanism, applicants who let an LLM rewrite their summary in that model's preferred register gain a systematic edge when that same model screens — a real and policy-relevant finding about AI-AI interaction that prior demographic-fairness work overlooked.