Surveilling Suitability: How AI Hiring Interviews Impact Job Seekers with Disabilities
Vaishnav Kameswaran, Valentina Hong, Jazmin Clark, Yu Hou, Hal Daumé III, Katie Shilton · 2026 · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) · doi:10.1145/3772318.3791516
Summary
This CHI 2026 paper reports a qualitative study of how AI-driven hiring interview platforms — asynchronous video interview tools (e.g., HireVue) that use AI to score candidates on facial expressions, vocal cues, and behavioural data — are perceived and experienced by job seekers with disabilities. The authors conducted five focus groups (11 participants) and eight semi-structured interviews, for a total of 19 participants with a broad range of disabilities including mobility, motor, blind/low vision, neurodivergent, learning, speech, chronic health, and mental-health conditions. Nine participants had direct experience completing AI hiring interviews. The 45-90 minute online sessions covered prior hiring experiences, fairness perceptions, disability disclosure strategies, and an 'incremental reveal' probe in which participants viewed HireVue marketing content followed by progressive explanations of how the AI scores candidates and infers disability signals. Transcripts were analysed through thematic analysis with 16 initial codes, organised around four overarching themes. The authors deploy workplace-surveillance theory (Lyon's 'surveillance as sorting,' Saltes' 'disability surveillance infrastructures') as the analytical frame, arguing that AI hiring interviews reconfigure the power relation between applicants and employers by collapsing a formerly two-way conversation into a one-way, algorithmically-gatekept assessment, with particular harms for disabled job seekers who are already underrepresented in the workforce.
Key findings
Participants consistently viewed AI hiring interviews negatively, describing them as 'messed up,' 'creepy,' 'dehumanizing,' and 'weird.' Four harm patterns emerged. (1) AI centres normative characteristics: systems codify a statistically-dominant 'ideal candidate' profile (steady eye contact, neurotypical affect, fluent voice), so participants with speech impairments, autism, ADHD, motor differences, or blindness felt penalised for non-normative behaviours and described intensive 'masking' labour to perform neurotypical/non-disabled presentation. (2) AI exacerbates information asymmetries: strict time limits and one-way recording remove the ability to ask clarifying questions, read interviewer reactions, or contextualise a disability as a strength — features participants relied on in traditional interviews to combat ableism. (3) AI undermines autonomy: automated disability-detection features (inferred from mouse movements, facial expressions, or speech patterns) take disclosure decisions away from candidates; several participants cited instances of successfully disclosing and negotiating accommodations in human interviews that would be impossible with AI. (4) AI intrudes on privacy: participants distrusted opaque data capture, feared their disability data would train future models or leak into HR records, and felt coerced into consenting because refusing meant no job. The authors frame these as 'surveillance infrastructures' that enact gatekeeping rather than disciplinary control, and offer design (community-led audits, participatory AI) and policy (ADA expansion, NYC-style automated-decision regulation covering disability) implications.
Relevance
For accessibility practitioners, HR leaders, procurement teams, and policymakers, this paper is a direct, empirically-grounded warning that current AI hiring products systematically disadvantage disabled candidates in ways that ADA-style reactive accommodation frameworks cannot address. The harms are structural, not bug-fixable: the platforms' one-way, normative-matching architecture is the problem, not individual model inaccuracies. Practical takeaways: (1) vendor due diligence for hiring AI must include disability-specific harms, not just race/gender fairness; (2) any deployment should preserve two-way interaction, let candidates ask questions, allow breaks and retakes, and disclose what signals are being analysed; (3) automated disability-detection features should be treated as presumptively prohibited, not opt-in; (4) community-led audits with disabled job seekers should gate procurement. Limitations: US-only sample of 19 recruited via disability advocacy groups and snowball sampling, which the authors acknowledge skews toward engaged advocates; the study does not include deaf participants as a distinct group, HR professionals, or Global South contexts where AI hiring adoption patterns differ.
Tags: disability · AI hiring · surveillance · algorithmic bias · employment · AI fairness · privacy · disability disclosure · workplace accessibility · automated video interview
Standards referenced: ADA · Americans with Disabilities Act