Not yet human-reviewed
This source record was written automatically by the research pipeline and has not yet been reviewed by the maintainer or integrated into a topic page. Treat it as a provisional summary and verify claims against the original source.
Computer says ‘no’: Exploring systemic bias in ChatGPT using an audit approach
Citation. Lippens, L. (2024). Computer says ‘no’: Exploring systemic bias in ChatGPT using an audit approach. Computers in Human Behavior: Artificial Humans, 2(1), Article 100054. https://doi.org/10.1016/j.chbah.2024.100054
Summary. This correspondence audit tests whether ChatGPT’s ratings of otherwise fictitious job applicants vary with names signalling ethnicity and gender. Across 34,560 vacancy–CV combinations, the study finds that both identity cues and job characteristics can affect ratings, with ethnic differences more pronounced than gender differences. The study frames these results as a risk for language-model use in personnel selection rather than evidence that a model’s ratings are appropriate for consequential employment decisions.
Key claims
- In the simulated CV-screening task, ChatGPT’s evaluations differed by applicant names signalling ethnicity and gender across 34,560 vacancy–CV combinations.
- Reported ethnic discrimination was more pronounced in jobs with favourable labour conditions or greater language-proficiency requirements.
- Reported gender bias was concentrated in gender-atypical roles, while some apparent ethnicity-by-gender penalties were not statistically significant after Holm correction.
Evidence & limitations
- The study uses a correspondence-audit design with fictitious applicant profiles and a GPT-3.5 snapshot dated 13 June 2023, so its estimates concern that simulated screening setup rather than live hiring outcomes or later model versions.
- The Drive artifact is a 39-page public arXiv PDF (arXiv:2309.07664) retrieved after the submitted Jina markdown was found to be a DOI 404 stub; Crossref and Unpaywall identify the canonical peer-reviewed article and DOI, but binary equivalence to the publisher version was not claimed.