Every AI hiring tool is trained, tuned, and validated against some implicit model of what a "normal" candidate looks like, sounds like, and moves like. For most disabled candidates, that model doesn't include them — and the result isn't always overt discrimination in the way people picture it. More often, it's a system that simply can't parse a valid, qualified response correctly, and scores the mismatch as a deficiency in the candidate rather than a limitation in the tool.
Where the Gaps Actually Show Up
Speech-to-text engines underlying video and voice interview scoring are trained predominantly on typical speech patterns. Candidates with speech-affecting conditions — stutters, dysarthria, conditions affecting speech clarity or pace — can be transcribed inaccurately, and inaccurate transcription cascades directly into inaccurate content scoring, since most systems evaluate what the candidate is understood to have said, not what they actually said. This is functionally similar to the accent-related transcription problem we flagged in our guide to AI phone screening, but the accommodation obligations are legally distinct — accent-related accuracy gaps are a fairness and bias issue; speech-disability-related gaps trigger ADA reasonable accommodation obligations directly.
Vision-based scoring — eye contact tracking, facial expression analysis, gesture and posture assessment — creates comparable problems for candidates who are blind or low-vision, have conditions affecting eye movement or facial muscle control, or use assistive technology that changes how they appear on camera. We noted in our guide to what AI interviews actually score that most major platforms, including HireVue, have moved away from facial-expression scoring specifically — partly in response to accuracy and bias concerns that overlap significantly with the disability-accessibility problem, even though the two aren't identical.
Timed assessments, common in the technical-screening category we reviewed in our Codility review, disadvantage candidates with certain motor, cognitive, or attention-related conditions unless accommodations — extended time, alternative input methods — are built in and easy to request. Codility's WCAG 2.2 investment, which we noted as a genuine differentiator, is a rare example of a vendor treating this as a first-class design requirement rather than an afterthought.
Neurodivergence: The Least-Discussed Gap
Autistic candidates and candidates with ADHD face a particular version of this problem that doesn't map neatly onto traditional accessibility categories like screen-reader compatibility. Many AI interview scoring models — especially older ones that weighted eye contact, vocal tone consistency, and conversational "naturalness" — penalize communication styles that are simply different rather than deficient: reduced eye contact, flatter vocal affect, more literal or direct answers, longer processing pauses before responding. None of these correlate with job performance, but several correlate strongly with autism and ADHD presentation, meaning a poorly designed scoring model can systematically screen out neurodivergent candidates while appearing, on the surface, to be measuring something neutral like "communication skills."
This connects directly to the pattern we identified in our AI hiring bias research summary: proxy variables that correlate with a protected characteristic without naming it directly are exactly the mechanism through which algorithmic discrimination tends to occur, and neurodivergent communication style is one of the clearest examples in the AI interview context specifically.
Most disability-related accessibility failures in AI hiring tools aren't intentional discrimination — they're the predictable result of building and validating a model against a narrow, unstated definition of "typical" candidate behavior, then deploying it at scale without testing it against a genuinely representative range of communication styles and abilities.
What the Law Requires
In the US, the Americans with Disabilities Act requires reasonable accommodation in the hiring process, and the EEOC has issued specific guidance stating that AI hiring tools do not exempt employers from this obligation — if an AI tool screens out a candidate because of a disability-related trait, and the employer didn't provide an accommodation or an equally effective alternative process, that can constitute disability discrimination under the ADA regardless of whether the discrimination was intentional. The EEOC's position is explicit that "the algorithm didn't mean to do it" is not a defense.
In the EU, the European Accessibility Act and the accessibility provisions woven into the broader digital services framework impose parallel obligations, and as we covered in our explainer on the EU AI Act and high-risk hiring classification, the AI Act's human oversight and Fundamental Rights Impact Assessment requirements for high-risk hiring systems specifically contemplate disability as a protected characteristic requiring active consideration, not passive compliance.
What Candidates Can Do
- Request accommodation before the interview, not during it — extended time, an alternative format, or a human-conducted interview in place of an AI one are all reasonable requests under the ADA, and making the request in writing creates a record.
- You are not required to disclose your specific diagnosis to request an accommodation — describing the functional limitation and the accommodation needed is legally sufficient in most US jurisdictions.
- Ask specifically whether the tool has been tested for accessibility — WCAG 2.2 conformance, screen-reader compatibility, and disability-inclusive validation testing are all fair, specific questions to put to a recruiter.
- If you believe you were screened out due to a disability-related trait, the EEOC's guidance on AI and the ADA is a useful reference point, and consulting an employment attorney is worth considering if the pattern seems clear.
What Employers Should Actually Check
- Ask vendors directly for their WCAG conformance level and any independent accessibility audit results — not marketing claims, actual documentation.
- Confirm an accommodation request pathway exists and is genuinely fast — a request that takes two weeks to process defeats the purpose.
- Ask specifically whether the scoring model weights eye contact, vocal tone, or "naturalness" as inputs, and if so, request evidence it's been validated against neurodivergent communication patterns.
- Treat accessibility testing as part of the same due diligence process as the bias auditing we described in our bias research summary — the two issues overlap more than most procurement processes currently acknowledge.
An AI interview tool that's never been tested against a blind candidate, a candidate who stutters, or an autistic candidate isn't accessible by omission — it's untested, which in a hiring context is functionally the same as being built for one kind of candidate and not the others.
This is the eighteenth and final piece in this phase of our AI recruitment research series. This piece describes general legal principles as of July 2026 and is not legal advice. Sharingan AI evaluates recruitment technology and policy independently, without vendor sponsorships or affiliate relationships.