Selection-Rate Ratios for Deployed AI Hiring Tools by Protected Group

confidence: high · 1 sources

What the evidence says

The claims on this page do not measure selection-rate ratios for deployed AI hiring tools. All six come from one study, Hofeditz (2022): an online experiment with a simulated AI recommendation system, a 2×2 between-subjects design, and N=194. Participants chose candidates with or without the system's recommendations. The outcome is the number of candidates selected from a group, not a ratio of selection rates, and no real deployed tool was tested. For the topic as named, the evidence here is absent.

The experiment does measure how recommendations changed which candidates people selected. For race, participants with recommendations selected fewer foreign-race candidates (M = 1.98) than those without (M = 2.11), a decrease of 0.13 candidates. The difference was not statistically distinguishable from zero (p = 0.472, r = 0.075). This is a no-measured-effect result.

For age and gender, recommendations increased selection. Participants with recommendations selected more older candidates (M = 2.33 against M = 1.91, an increase of 0.42, p = 0.012, r = 0.251). They also selected more female candidates (M = 1.78 against M = 1.45, an increase of 0.33, p = 0.03, r = 0.214). The source labels both effect sizes medium. The recommendations were simulated, so the magnitudes may not carry over to deployed systems. The study states no mechanism for why recommendations shifted selections toward older or female candidates.

The study also tested whether adding explainable AI (XAI) changed these effects. It increased the explained variance in foreign-race selection by 4.66% (p = 0.002, 95% CI [0.279, 1.236]). It showed no measured moderation for older candidates (p = 0.772) or for female candidates (p = 0.444). The source says XAI moderated the effects differently depending on context.

Everything rests on a single experiment. Nothing here covers other protected groups, other tools, or real hiring outcomes.

Where the evidence splits

No contradictions in the current sources.

Thresholds and dose-response

No threshold established in the current sources.

Open empirical questions

What are the selection-rate ratios, by protected group, for AI hiring tools actually deployed in employment settings? The sources here measure none.
Do the age and gender increases found with a simulated recommender appear when real tools and real applicant pools are used?
Why did XAI moderate the foreign-race effect but not the age or gender effects?
What happens to selection of other protected groups, such as disability status? The sources here test only race, age, and gender.

Findings

AI recommendations show no measured effect on selection of foreign-race candidates

confidence — direction: mediumconfidence — magnitude: lowconsensus: thin — a single study.

mechanism: not established. Participants chose candidates with or without simulated AI recommendations.
effect: no measured effect. Selection fell numerically by 0.13 candidates (M = 1.98 with recommendations, M = 2.11 without). The difference was nonsignificant (p = 0.472, r = 0.075).
conditions / moderators: XAI moderated this effect (see the XAI finding below). Not established beyond one simulated-system experiment.

Lennart Hofeditz (2022), Electronic Markets · doi:10.1007/s12525-022-00600-9 · https://link.springer.com/content/pdf/10.1007/s12525-022-00600-9.pdf (retrieved 2026-10-01)

AI recommendations increase selection of older candidates

confidence — direction: highconfidence — magnitude: mediumconsensus: thin — a single study.

mechanism: not established.
effect: increases selection by 0.42 candidates (M = 2.33 with recommendations, M = 1.91 without; p = 0.012, r = 0.251, medium).
conditions / moderators: simulated AI system in an online experiment, not a deployed tool. XAI showed no measured moderation (p = 0.772).

Lennart Hofeditz (2022), Electronic Markets · doi:10.1007/s12525-022-00600-9 · https://link.springer.com/content/pdf/10.1007/s12525-022-00600-9.pdf (retrieved 2026-10-01)

AI recommendations increase selection of female candidates

confidence — direction: highconfidence — magnitude: mediumconsensus: thin — a single study.

mechanism: not established.
effect: increases selection by 0.33 candidates (M = 1.78 with recommendations, M = 1.45 without; p = 0.03, r = 0.214, medium).
conditions / moderators: simulated AI system in an online experiment, not a deployed tool. XAI showed no measured moderation (p = 0.444). The Levene test showed unequal variances, so the Welch test was used.

Lennart Hofeditz (2022), Electronic Markets · doi:10.1007/s12525-022-00600-9 · https://link.springer.com/content/pdf/10.1007/s12525-022-00600-9.pdf (retrieved 2026-10-01)

Explainable AI moderates the recommendation effect on foreign-race selection, but not on age or gender selection

confidence — direction: highconfidence — magnitude: medium for race, high for age and genderconsensus: thin — a single study.

mechanism: not established. The source reports that XAI moderated the effects differently depending on context.
effect: for foreign-race selection, the moderation explains 4.66% additional variance (F(1, 190) = 9.33, p = 0.002, 95% CI [0.279, 1.236]). For older candidates there is no measured moderation (ΔR2 < 0.01%, p = 0.772, 95% CI [-0.518, 0.384]). For female candidates there is no measured moderation (ΔR2 < 0.01%, p = 0.444, 95% CI [-0.276, 0.613]).
conditions / moderators: moderation analysis with 5,000 bootstrap iterations in one online experiment.

Lennart Hofeditz (2022), Electronic Markets · doi:10.1007/s12525-022-00600-9 · https://link.springer.com/content/pdf/10.1007/s12525-022-00600-9.pdf (retrieved 2026-10-01)