In controlled tests, resume‑screening LLMs preferred resumes generated by themselves over equally qualified human‑written or other‑model resumes. Self‑preference bias ran 68%–88% across major models, boosting shortlists 23%–60% for applicants who used the same LLM as the evaluator. Simple prompts/filters halved the bias.
— This reveals a hidden source of AI hiring unfairness and an arms race incentive to match the employer’s model, pushing regulators and firms to standardize or neutralize screening systems.
msmash
2026.01.16
70% relevant
The Code.org hiring ban engages the same public‑interest domain as demonstrated AI bias in hiring: the story raises the practical question of how employers should regulate candidate use of generative AI when such tools can distort evaluation and create arms‑race incentives for applicants to tune outputs to the employer’s stack.
msmash
2026.01.14
85% relevant
The Guardian/CaseBasix report says McKinsey asks candidates to prompt, review and apply judgement to outputs from its internal assistant (Lilli). That ties directly to the existing finding that LLM‑based screening and evaluation systems can prefer candidates who match the evaluator model’s style or who are skilled at prompt‑engineering, creating an arms‑race incentive for applicants to optimize for the employer’s model rather than underlying competence.
Arctotherium
2025.12.03
72% relevant
Both pieces document evaluative biases in LLMs: the existing idea shows models prefer resumes produced by themselves, while the article reports models assigning higher moral value to people with particular political identities (environmentalists/socialists/communists). Together they point to a broader problem of LLMs producing systematic, non‑transparent preference skews that can distort hiring, moral judgments, and other downstream decisions (the article names Claude as an actor exhibiting a strong preference).
Tyler Cowen
2025.10.03
100% relevant
Paper by Jiannan Xu, Gujie Li, and Jane Yi Jiant reporting self‑preference bias and mitigation in resume screening across 24 occupations.