What the evidence says
The strongest measured evidence on this page is one controlled experiment on age. Rašticová (2026) ran a 2×2 factorial vignette study with between-subjects manipulation of age (32 vs 56) and gender. Candidate qualifications were equivalent. Older candidates were rated lower on competence-related traits (F(1, 604) = 17.09, p < 0.001, ηp2 = 0.028) and lower on emotional stability, optimism and resilience (F(1, 604) = 25.90, p < 0.001, ηp2 = 0.041). The effect sizes are small to medium. The study does not involve an AI system. It measures the human perception of candidates that an age-biased tool could inherit or reproduce.
The same study finds that age and gender interact. Younger women (F32) were perceived as having smaller digital skill gaps than F56, M32 and M56 profiles (interaction p = 0.002). Younger candidates, especially younger women, also received more favorable ratings on interpersonal contact skills (age effect p = 0.012).
For race, the evidence is associational or descriptive. Hunkenschroer (2022) reports that recruiter ratings of applicants' Facebook information were unrelated to later job performance and produced subgroup differences favoring female and Caucasian applicants (van Iddekinge et al., 2016). The magnitude is not quantified. The review also reports that several studies found facial recognition software has disparities in error rates across gender and race, again without a magnitude. It states that image and voice recognition techniques can predict applicants' race, age and sexual orientation, and that other algorithms may forecast pregnancy likelihood. These claims show that a system can infer protected traits. They do not show that hiring decisions change as a result.
Two further claims are mechanisms raised as concerns, not measured effects. Critics argue that algorithms trained on current-employee data may discriminate against underrepresented groups if employees do not represent the applicant pool. Fernández-Martínez and Fernández (2020) warned that emotion recognition software may disadvantage specific races or ethnic groups, because it may not account for differences in intonation or emotional expression across languages and cultures.
One audit of a single system finds no disparity on the axes it tested. Lee (2026) reports performance within ±3% across junior, mid-level and senior groups, a Statistical Parity Difference below 0.05, a Disparate Impact Ratio of 0.85 to 1.12, and no systematic bias amplification. The axes are proxy variables such as experience level, not actual demographic attributes. The author notes that bias conclusions are incomplete without demographic data. The audit also reports a minor skew in niche skill domains, attributed to data imbalances.
Coverage is thin for several expected axes. Sánchez-Monedero (2020) finds that the automated hiring systems it reviewed do not test social class or disability. They test separate demographic axes and do not directly compare combinations, such as Hispanic women with white men. Research on intersectional and rich-subgroup fairness in machine learning is described as limited and early-stage.
Findings
Older candidates are rated lower on competence-related traits than younger candidates with equivalent qualifications
confidence — direction: highconfidence — magnitude: mediumconsensus: thin — a single experiment
- mechanism
- not established; the study manipulates candidate age in a vignette and measures rater perception
- effect
- decreases ratings; F(1, 604) = 17.09, p < 0.001, ηp2 = 0.028, on a 10-point scale across 10 traits
- conditions / moderators
- equivalent qualifications; ages 32 vs 56; human raters, not an AI system
Older candidates are rated lower on emotional resilience and optimism than younger candidates with equivalent qualifications
confidence — direction: highconfidence — magnitude: mediumconsensus: thin — a single experiment
- mechanism
- not established
- effect
- decreases ratings; F(1, 604) = 25.90, p < 0.001, ηp2 = 0.041
- conditions / moderators
- equivalent qualifications; the effect size is larger than for the competence dimension
Age and gender interact in perceived digital skill gaps and interpersonal contact skills, with younger women rated most favorably
confidence — direction: highconfidence — magnitude: mediumconsensus: thin — a single experiment
- mechanism
- not established
- effect
- F32 profiles were perceived as having smaller digital skill gaps than F56, M32 and M56 (interaction p = 0.002, Tukey post-hoc). Younger candidates received more favorable interpersonal contact ratings (age effect p = 0.012).
- conditions / moderators
- 2×2 vignette design; not tested beyond two ages and two genders
One audited hiring system shows no measured disparity across proxy subgroups
confidence — direction: mediumconfidence — magnitude: mediumconsensus: thin — a single audit on proxy axes
- mechanism
- not applicable; post hoc fairness metrics and auditing
- effect
- no measured effect. Performance varied within ±3% across experience levels, SPD was below 0.05, DIR ranged from 0.85 to 1.12, and there was no systematic bias amplification. A minor skew appeared in niche skill domains.
- conditions / moderators
- proxy variables only, not actual demographic attributes; the author says bias conclusions are incomplete without demographic data
Algorithmic assessment of social media data produces subgroup differences favoring Caucasian applicants
confidence — direction: highconfidence — magnitude: mediumconsensus: thin — one cited study
- mechanism
- not established; the source reports that ratings were unrelated to later job performance
- effect
- increases ratings for Caucasian (and female) applicants; magnitude not quantified
- conditions / moderators
- the cited study concerns recruiter ratings of Facebook information
Facial recognition software shows disparities in error rates across racial groups
confidence — direction: highconfidence — magnitude: mediumconsensus: thin — reported by the review from several studies, none quantified here
- mechanism
- not established in the claim
- effect
- increases error-rate disparities across gender and race; magnitude not quantified in this review
- conditions / moderators
- facial recognition software generally, not a hiring outcome
Image, voice and prediction techniques can infer applicants' race, age, sexual orientation and pregnancy likelihood
confidence — direction: highconfidence — magnitude: mediumconsensus: thin
- mechanism
- models infer protected traits from image, voice or other applicant data
- effect
- enables inference of the traits; accuracy and error rates not provided
- conditions / moderators
- the source uses hedged language for pregnancy forecasting
Training on current-employee data and emotion recognition software are proposed pathways to racial and group disadvantage
confidence — direction: mediumconfidence — magnitude: lowconsensus: thin — stated as concerns
- mechanism
- a biased training sample that does not represent the applicant pool; emotion recognition that ignores differences in intonation and expression across languages and cultures
- effect
- possible disadvantage to underrepresented groups and specific races or ethnic groups; not measured
- conditions / moderators
- both claims are concerns or warnings, not observed effects
Reviewed hiring systems do not test social class, disability or intersectional combinations
confidence — direction: highconfidence — magnitude: highconsensus: thin
- mechanism
- not applicable; a gap in documented validation
- effect
- no measured effect, because these axes are not tested
- conditions / moderators
- based on review of publicly available system documentation and patents; research on intersectional fairness in ML is limited and early-stage