Disability and Age as Axes in Algorithmic Hiring Audits

confidence: high · 4 sources

What the evidence says

The sources on this page are mostly reviews and conceptual papers. They flag disability and age as risk axes in algorithmic hiring. They rarely measure the size of any effect. Field experiments, as summarized in one survey, "point to a risk of direct discrimination against disabled and older applicants." That summary gives no rates, so direction is medium-confidence and magnitude is unquantified.

The sources name several routes by which disadvantage could arise. Accessibility problems and ableist norms can discourage disabled people from applying at all. Asynchronous video interviews can produce patterns from candidates with speech impairment, such as short answers. They can also mismeasure input from candidates with sight impairment, such as eye contact. Algorithmic recruitment models judge both unfavorably. On age, advertising campaigns run on platforms that cater to younger users are less likely to reach older people. None of these routes comes with a measured effect size on this page.

One source argues that discrimination against people with disabilities (PWDs) differs in kind from discrimination against other protected groups. It calls it mainly heterogeneous. For example, an applicant who uses assistive technology may score lower on timing metrics. A candidate with autism may avoid eye gaze, which a system may misread as low confidence. This is illustrated by examples, not measured. The same source says removing sensitive data from a model's input ("fairness through unawareness") is insufficient. Sensitive information can remain correlated with other features or proxies. This is a cited result from fairness research, not a new measurement.

The sources also show that disability is often left out of fairness work. In a study of three automated hiring systems, bias was actively reduced for some protected traits. Disability was not mentioned in the examples or the validation tests. In a keyword search of 82 AI ethics guidelines, 22 (26.8%) mentioned disabled people as a vulnerable group. For context, one source cites European unemployment rates for people aged 20–64 of 17.1% with disabilities and 10.2% without. That is a gap of 6.9 percentage points. This is a labor-market statistic, not a measure of algorithmic hiring.

Evidence on outcomes for vulnerable populations is thin. One paper cites Abid et al. (2024), who find that AI's employment effect on a vulnerable population reverses from negative to positive beyond critical adoption thresholds. The paper itself does not validate this. It reads the result as showing that outcomes depend on implementation design, not on whether AI is present. The same paper asserts that transparency, fairness and accessibility in AI competency-assessment tools are necessary to prevent algorithmic discrimination. It offers no measurement of that relationship.

Vendor material is a separate kind of evidence. Retorio and Interviewer.ai claim their systems evaluate candidates regardless of age, or are blind to age. Retorio describes comparing mean Big 5 trait estimates across young and old people. These are marketing and vendor-reported statements. The source reviewing them does not test whether the systems behave as claimed. No source here measures age-based score differences, selection rates or callback gaps in an algorithmic hiring system.

Where the evidence splits

No contradictions in the current sources.

Thresholds and dose-response

No threshold established in the current sources.

Open empirical questions

How large are the score or selection-rate gaps by disability status and by disability type in deployed algorithmic hiring systems?
How large are the score or selection-rate gaps by age in deployed hiring systems, and do vendor claims of age-blindness hold under independent audit?
By how much do inaccessible application processes reduce application rates among disabled people?
At what adoption level does the reported reversal in AI's employment effect occur, and does it apply to disabled or older workers specifically?

Findings

Field experiments indicate a risk of direct discrimination against disabled and older applicants

confidence — direction: mediumconfidence — magnitude: lowconsensus: thin — one survey's summary of experiments, no figures

mechanism: not established in the claim; the source states a risk of direct discrimination
effect: direction is toward worse treatment of disabled and older applicants; magnitude is not quantified
conditions / moderators: not established; the claim does not give rates or the experiments' settings

Alessandro Fabris (2023), arXiv · doi:10.1145/3696457 · https://arxiv.org/pdf/2309.13933v4 (retrieved 2026-10-02)

Accessibility issues and ableist norms decrease job application likelihood among disabled people

confidence — direction: mediumconfidence — magnitude: lowconsensus: thin

mechanism: inaccessible application processes and ableist norms discourage people from applying
effect: decreases applications from disabled people; magnitude not quantified
conditions / moderators: not established

Alessandro Fabris (2023), arXiv · doi:10.1145/3696457 · https://arxiv.org/pdf/2309.13933v4 (retrieved 2026-10-02)

Asynchronous video interviews decrease algorithmic evaluations of candidates with speech or sight impairments

confidence — direction: mediumconfidence — magnitude: lowconsensus: thin

mechanism: speech-impairment patterns such as short answers, and mismeasured eye contact for sight impairment, are judged unfavorably by algorithmic recruitment models
effect: decreases evaluation scores for these candidates; magnitude not quantified
conditions / moderators: applies to asynchronous video interviews; other formats not established

Alessandro Fabris (2023), arXiv · doi:10.1145/3696457 · https://arxiv.org/pdf/2309.13933v4 (retrieved 2026-10-02)

Job advertising on platforms that cater to younger users decreases reach to older job seekers

confidence — direction: mediumconfidence — magnitude: lowconsensus: thin

mechanism: platform audience composition means campaigns on youth-oriented platforms reach fewer older people
effect: decreases reach to older segments; magnitude not quantified
conditions / moderators: applies to campaigns run on platforms that cater to younger users

Alessandro Fabris (2023), arXiv · doi:10.1145/3696457 · https://arxiv.org/pdf/2309.13933v4 (retrieved 2026-10-02)

Disability is omitted from bias validation while other protected traits receive bias reduction

confidence — direction: highconfidence — magnitude: highconsensus: thin — one cited study of three systems

mechanism: validation tests and examples cover some protected traits and leave disability out, so disability-related bias goes unchecked
effect: in three automated hiring systems studied, bias was actively reduced for some traits and disability was not mentioned in the examples or validation tests
conditions / moderators: observed in three systems; the source cites the study and does not run it

Maarten Buyl (2022), arXiv · doi:10.1145/3531146.3533169 · https://arxiv.org/pdf/2206.06149v1 (retrieved 2026-10-02)

AI ethics guidelines rarely name disabled people as a vulnerable group

confidence — direction: highconfidence — magnitude: highconsensus: thin — one count

mechanism: not applicable; this is a count of document content
effect: 22 of 82 guidelines (26.8%) mentioned disabled people or people with disabilities as a vulnerable group deserving special consideration
conditions / moderators: keyword search for "disab*" across the guidelines' documents and websites

Maarten Buyl (2022), arXiv · doi:10.1145/3531146.3533169 · https://arxiv.org/pdf/2206.06149v1 (retrieved 2026-10-02)

Discrimination against people with disabilities in automated hiring is heterogeneous

confidence — direction: mediumconfidence — magnitude: mediumconsensus: thin

mechanism: different disabilities produce different mismatches with system metrics, such as assistive-technology users scoring lower on timing or autistic candidates avoiding eye gaze
effect: requires disability-specific assessment; no magnitude given
conditions / moderators: the source contrasts this with traditionally considered protected groups; shown by examples, not quantitative evidence

Maarten Buyl (2022), arXiv · doi:10.1145/3531146.3533169 · https://arxiv.org/pdf/2206.06149v1 (retrieved 2026-10-02)

Removing disability data from model input has no measured effect on preventing discrimination

confidence — direction: mediumconfidence — magnitude: mediumconsensus: thin — a cited background result

mechanism: sensitive information stays correlated with other, non-protected features or proxies
effect: removal alone is described as insufficient; no magnitude given
conditions / moderators: cited from fairness research in general, not tested on hiring systems in this source

Maarten Buyl (2022), arXiv · doi:10.1145/3531146.3533169 · https://arxiv.org/pdf/2206.06149v1 (retrieved 2026-10-02)

European unemployment is higher for people with disabilities than for people without

confidence — direction: highconfidence — magnitude: highconsensus: settled as a descriptive gap; its link to algorithmic hiring is not established here

mechanism: not established in the claim
effect: 17.1% unemployment for people with disabilities aged 20–64 versus 10.2% for people without, a gap of 6.9 percentage points
conditions / moderators: Europe, ages 20–64; cross-sectional figure that does not isolate algorithmic hiring

Maarten Buyl (2022), arXiv · doi:10.1145/3531146.3533169 · https://arxiv.org/pdf/2206.06149v1 (retrieved 2026-10-02)

Employment outcomes of AI for a vulnerable population reverse from negative to positive depending on adoption level

confidence — direction: highconfidence — magnitude: mediumconsensus: thin — one cited study

mechanism: outcomes depend on implementation design, according to the citing paper
effect: mixed evidence; the effect reverses from negative to positive beyond critical adoption thresholds, with no threshold value given
conditions / moderators: the source cites Abid et al. (2024) and does not validate it; the population is described only as vulnerable

Roh BG. (2026), Frontiers in sociology · doi:10.3389/fsoc.2026.1891969 · https://europepmc.org/articles/PMC13549850?pdf=render (retrieved 2026-10-02)

Transparency, fairness and accessibility in assessment tools are asserted as necessary to prevent algorithmic discrimination

confidence — direction: highconfidence — magnitude: lowconsensus: thin — an assertion with a citation to prior work

mechanism: not measured; the source states necessity as a normative design claim
effect: no measured effect; the relationship is asserted, not tested
conditions / moderators: not established

Roh BG. (2026), Frontiers in sociology · doi:10.3389/fsoc.2026.1891969 · https://europepmc.org/articles/PMC13549850?pdf=render (retrieved 2026-10-02)

AI hiring vendors state that their tools ignore age; the sources do not test this

confidence — direction: highconfidence — magnitude: lowconsensus: thin — marketing claims only

mechanism: vendors describe age-blind scoring, and one describes comparing trait estimates across age groups in its training data
effect: no measured effect; these are vendor claims of neutrality, not outcome measurements
conditions / moderators: whether the systems behave as claimed is not tested in the source

Drage E (2022), Philosophy & technology · doi:10.1007/s13347-022-00543-1 · https://europepmc.org/articles/PMC9550152?pdf=render (retrieved 2026-10-03)