TRV-2026-3027Version 1 · Published

Saved · see the current record

Why this version exists

Certified into the record

The exact text we saved

This is the version exactly as it was saved, character for character. The SHA-256 hash below is a 64-character code worked out from that text. Run the same text through SHA-256 and you get the same code; change a single letter and the code comes out completely different. Each version also stores the previous version’s hash, which chains them together so no version can be swapped out quietly.

TRUVACE RECORD VERSION
record: TRV-2026-3027
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-18T12:02:26.629919Z
status: published
lens: p_space
sector: labor
headline: Automating inequality? Gender bias in AI-mediated labor markets
dek: Generative AI is rapidly entering labor markets, with large language models (LLMs) now used to help write CVs and evaluate applicants. We study the risks this creates in a series of audit experiments in which applicant gender is randomly assigned while all other information is held constant. We find that LLMs generate gender-stereotypical CVs and exhibit gender bias in screening scores and wage recommendations. Screening scores track occupational gender segregation, penalizing women in male-dominated occupations…
gain_title: (none)
problem_title: In audit experiments with randomly assigned applicant gender, large language models wrote gender-stereotypical CVs, penalized women in male-dominated occupations in screening scores and favored them in female-dominated ones.
trace_subject: (none)
gain_reading: (none)
gain_evidence: (none)
problem_reading: In audit experiments with randomly assigned applicant gender, large language models wrote gender-stereotypical CVs, penalized women in male-dominated occupations in screening scores and favored them in female-dominated ones.
problem_evidence: LLMs generate gender-stereotypical CVs and exhibit gender bias in screening scores and wage recommendations | penalizing women in male-dominated occupations and favoring them in female-dominated ones
quick_read: A study in PNAS ran audit experiments in which applicant gender was randomly assigned while all other information stayed the same. Large language models produced gender-stereotypical CVs and gave different screening scores and wage recommendations depending on gender.

Screening scores followed the gender makeup of occupations: women were penalized in male-dominated jobs and favored in female-dominated ones. Wage recommendations followed occupational gender wage gaps when gender was stated outright in the prompt.

The authors describe two routes to the disparities, and say that because the same systems both write CV content and assess it, the routes can reinforce each other. They describe the result as measurable risks of algorithmic amplification in AI-mediated job search.
limitation: The wage-recommendation pattern weakened when gender was signaled by names rather than stated explicitly, and it was sensitive to which language models were used.
tag: Evidence-backed problem
key_points: Researchers used audit experiments in which applicant gender was randomly assigned and all other information was held constant. | The language models generated gender-stereotypical CVs. | Screening scores penalized women in male-dominated occupations and favored them in female-dominated occupations. | Wage recommendations tracked occupational gender wage gaps when gender was explicit in the prompt. | The wage gradient weakened when gender was signaled only through names and varied depending on which model was used. | Evaluation bias means identical CV content is judged differently by gender; generation bias means gendered content the model produced affects later evaluation.
rundown: The study tested how large language models behave in two roles now common in hiring: helping write CVs and evaluating applicants. Randomly assigning gender while keeping everything else fixed let the researchers attribute differences in outcomes to gender.

Screening scores mirrored occupational gender segregation. Women received lower scores in male-dominated occupations and higher scores in female-dominated ones. Wage recommendations mirrored occupational wage gaps when gender was stated explicitly.

The authors identify two channels: evaluation bias, where the same CV content is rated differently by gender, and generation bias, where the model's own gendered CV content shapes later evaluation. Since one system can shape both CV content and its assessment, the two channels can compound.
sources:
- peer_reviewed | Proceedings of the National Academy of Sciences | https://doi.org/10.1073/pnas.2534118123 | 2026-11-16
prev: 0000000000000000000000000000000000000000000000000000000000000000
SHA-256 hash
f198c23d577028d305dfc6b6b9068393aab4f0659b55702cc56c36b784724ed0
Previous version’s hash
0000000000000000000000000000000000000000000000000000000000000000
Check this record hasn’t been altered
How to check it yourself, without trusting this page

Download any version from /api/record/TRV-2026-3027 and hash its canonical field yourself, for example with shasum -a 256. You should get its content_hash. Each version’s text also ends with prev:and the hash of the version before it (version 1 points to 64 zeros). If anyone had changed a single character of any version after it was saved, those hashes wouldn’t line up.