What the Record Shows About AI Interviews, Accents, and Disability

What These Tools Actually Score

Mostly the words, not the voice or the face. At least one major vendor states plainly that its AI-scored interviews run only on the transcribed text of an answer: no facial expression, no body language, no tone of voice, and no background or surroundings ¹. That single fact moves the fear most people bring to a one-way video interview: away from how you looked or sounded, and toward what the transcript of what you actually said contains.

This is one vendor's own description of one product's current configuration, not a claim about the whole market, and it does not mean the video goes unwatched: the recording still exists, and a recruiter can still open it. Check whether the interview shows a consent screen before you start, and read what it says the assessment evaluates. If there is no screen, or it says nothing about scoring, ask the recruiter directly what the tool reads.

Why a Transcript Is Not a Neutral Witness

Scoring only the words does not make the words neutral, because a machine transcript is not a direct copy of what you said. Five commercial speech-recognition systems tested on matched conversational interviews produced roughly double the word error rate for Black speakers as for white speakers, an average of 0.35 against 0.19 ². A wrong word in the transcript is what a scoring model reads as your answer, not a footnote next to it.

One interview vendor publishes its own contrasting numbers: internal testing puts its supplier's error rate under 10% on average, in a 7-10% range across the ethnicities it tested, against 15-30% for the alternative systems it compared ³. Read that as an internal, unreplicated comparison rather than an independent audit, since no sample size, corpus or methodology accompanies it, and hold it next to the published research rather than in place of it. For speech disabilities specifically, a 2025 research challenge built to close the gap still put its best system's error rate at roughly one word in twelve, and the organizers said plainly that recognition for disordered speech remains inadequate ⁴. Note what these two measurements cover and what they do not: speaker race in one, speech disability in the other. Accented and second-language speech is a separate literature, and no figure in this article measures it.

Format changes the size of a different problem in the other direction. A meta-analysis of first impressions of autistic people found the largest unfavorable gap in studies that showed video, and the smallest, with a confidence interval crossing zero, in studies that presented a transcript ⁵. Those were lab impressions rather than hiring decisions, and the transcript subgroup was the smallest one. Read the two findings together: a transcript-scored interview carries a measured transcription problem alongside a smaller presentation problem, which makes it neither neutral nor obviously the worse format.

Read the Documented Case Carefully

One filing gives this question something concrete. In March 2025 the ACLU of Colorado filed a complaint with the Colorado Civil Rights Division and the EEOC alleging that an AI-scored video assessment run on HireVue's technology produced a worse outcome for a deaf and Indigenous Intuit employee after her request for human-generated captioning was denied ⁶. HireVue disputed that its AI-based assessment was the tool actually used, and nothing in the complaint has been adjudicated.

Treat it as what it is: a documented allegation moving through a real civil-rights process, not a finding, and the clearest evidence available that the fear behind this question is not imaginary. It also sets a standard for the two answers that dominate a search on this subject, both of which are unfalsifiable. Vendors say their models are audited and do not read appearance, without publishing the audit; creator content says AI interviews judge your face, without naming a platform, a date or a document anyone can open. Ask of any claim here whether it points at a specific, checkable filing or published source, or only at a general impression of what these tools probably do.

Ask for a Different Format, Not a Different Verdict

If a barrier is about how you speak rather than what you know, ask for a different format before the interview, not an appeal after it. On at least one platform, a "Request Accommodations" button on the instructions screen relaxes time limits for a qualified disability, and the employer decides whether to grant the request ⁷. The same screen says the game-based section is not presented if you take that route, so ask what changes as well as what is relaxed.

Where the barrier is a speech disability, say so when you ask rather than describing the format on its own, and how to actually ask for one covers what to write and who to send it to. What an employer is told to offer in return is worth reading directly rather than guessing at: what an AI-based assessment owes a disabled applicant under the ADA covers the same US statute, the Americans with Disabilities Act of 1990, from the hiring side, including where a written or untimed alternative is the right fix and where it would change what the assessment is actually measuring. Disability law differs by country and by state, so check what applies where you are applying.

How to Check What a Specific Tool Reads

Ask the recruiter directly which of three things the tool scores: transcript content, delivery timing and pace, or video and audio characteristics. A well-run process can answer that in one sentence. Some platforms also put an AI consent statement on the screen before the assessment starts, and where one appears it is worth reading before you click through, since what it covers is set by the employer deploying the tool.

The same caution applies downstream of the interview, on a page you never see. Automated writing checks used elsewhere in an application carry their own measured accuracy gap for non-native English writers, covered separately here, so a written stage is not automatically the fairer one either. No format in this pipeline is free of a measured gap, which is why the question worth putting to a recruiter is what this particular tool reads.

See how it works >