What the evidence says
The sources here are thin on the exact question, which is outcome audits of deployed AI hiring systems that use actual protected demographic data. Almost all of it is secondhand: two review papers summarizing other studies and one audit of personality-scoring tools. None of the claims reports a full audit of real hiring outcomes broken out by protected group, and none gives a selection-rate ratio.
The clearest case is Amazon's recruiting algorithm, described in both reviews. One review says it ran in test mode. It assigned lower scores to women's resumes when ranking candidates. The other review says the system penalized resumes that mentioned women's colleges or used verbs more frequently used by female candidates. Both name the same mechanism. The algorithm was trained on data from current top performers, who were mostly male, and it learned to treat features associated with women as predictors of lower hiring likelihood. The claims give the direction but no score gap.
Two other findings concern tools at other stages of hiring. Facebook ads for supermarket cashier positions were shown to an audience that was 85% women, which suggests that sourcing algorithms can produce group disparities before any applicant is screened. Facial recognition software shows disparities in error rates across gender and race, according to three cited studies. The claim gives no rates. A separate line of work finds that commercial speech recognition systems from Apple, IBM, Google, Amazon, and Microsoft have higher word error rates for Black speakers than for White speakers on the same spoken content. These are not hiring audits, but they bear on any hiring tool that transcribes or analyzes speech.
The one audit with protected-group data tested personality scores on resumes submitted in different formats. The full sample stayed above a rank-order stability of 0.90 for Steadiness and Conscientiousness. The female subgroup (N = 33) and the English-primary-language subgroup (N = 56) fell below 0.90. These subgroups are small, and the claim does not say whether the gap is statistically meaningful.
One finding concerns human raters, not an algorithm. Recruiter ratings of applicants' Facebook information favored female and Caucasian applicants and were unrelated to later job performance. It is a useful comparison point, but it does not audit an AI system.
On regulation, New York City Local Law 144 requires third-party bias audits and public disclosure of the resulting ratios before automated employment decision tools are deployed. The EU AI Act classifies AI used in labor-related decision-making as high risk. The sources describe these rules but report no audit results produced under them.