When an AI Interview Asks Something It Should Not, Keep the Record

Which Questions Are Off Limits Before an Offer?

Very few by name, and the sharp exception is disability. Federal law bars an employer with 15 or more employees from asking a job applicant whether they have a disability, or about the nature or severity of one, before a conditional offer ¹. Questions about the ability to perform the job's functions are allowed ¹. That line sits in the ADA itself, at 42 U.S.C. 12112(d)(2).

The distinction is narrower than it sounds. "Can you do the lifting this route needs, with or without a reasonable accommodation" is a different question from "do you have a back problem". Only one of them belongs before an offer, and the rule governs what the employer may ask rather than what you may request.

Most of the rest of federal law draws its lines differently, and the sources here do not map them. What they do show is narrower: the pre-offer disability inquiry ban ¹, and the EEOC's position on fluency requirements ². The employer-side version of this ground asks what has to be offered rather than what may be asked: do ADA accommodations apply to an AI-based assessment.

A question that feels wrong is not automatically a violation.

National origin is where the line blurs, because the question can arrive dressed as a job requirement. The EEOC's 2016 enforcement guidance says an English fluency requirement is generally permissible only where the position needs it for effective performance ². Applying one uniform standard across a broad range of dissimilar positions may violate Title VII, the same guidance says.

That is guidance, not law, and it was written about stated requirements rather than about a machine transcript of your answer.

Should You Answer It, Skip It, or Say Why?

All three are legitimate, and the format gives you no way to tell in advance which one lands best. HireVue's candidate help center says every candidate for the same position gets the same employer-selected questions, with the same time to review and prepare ³. The question was configured for a role, not chosen for you, and there is no interviewer on the other end to read your face while you decide.

• Answer it. If the honest answer costs you nothing, answering costs you nothing, and it keeps the recording simple.
• Answer the part the job needs. Something like "what I can speak to is that I'm available for the shift pattern in the posting" gives the employer the information the role turns on and leaves the rest alone.
• Leave it and say why. "I'd rather not answer that one, and I'm glad to take anything else about the role" is a complete recording of a choice you made.

The honest limit is that you cannot see what happens next. You are not shown how a skipped question is weighed, or whether anyone notices the gap before a person opens the recording.

What you cannot do is redirect. In a live interview a good answer can move the conversation somewhere better, and a scripted prompt has no conversation to move. The timer runs, the next question loads, and whatever you said is what the employer receives.

None of the three is dishonest.

One thing worth deciding before you record: whether you still want the job. A question set that asks about family plans tells you something about the employer, and you are allowed to use that.

Withdrawing is a move too, and it is the easiest one to forget.

Why a Graduation Year Is the Same Question

Because it carries the answer without anyone asking for it. A high school graduation date implies an age. In a field experiment sending fictitious applications to entry-level jobs at 108 large United States employers, applications listing a high school graduation date implying an age over 40 drew 0.6 percentage points fewer employer contacts, with a cross-firm standard deviation of 1.1 points ⁴.

That penalty is small in absolute terms, the signal is indirect, and firm behavior varies enough that the average hides real differences between employers. It is evidence that a date can move employer contact, not evidence that any particular employer used yours.

Proxies are why a list of banned questions runs out early.

The clearest public example is a cutoff, not a model. The EEOC's description of the iTutorGroup settlement is a date-of-birth cutoff written into tutor application software, set to reject female applicants aged 55 or older and male applicants aged 60 or older ⁵. Three companies agreed to pay $365,000 to settle the EEOC's claims, which covered more than 200 qualified United States applicants the agency said were rejected because of their age ⁵. A consent decree resolves a case with no admission of liability.

Commentary calls it the first AI hiring settlement, and the agency describes automated screening.

Whether an employer would ever see the pattern in its own funnel is a separate question. The version hiring teams are given is how to run an adverse impact audit when the vendor holds the data.

Keep the Record Before You Close the Tab

Write down the exact wording, the date, the platform and the role, before anything closes. You may not get the recording. HireVue's candidate help center says a candidate has no direct access to review their own recording through the platform, and has to ask the hiring company, which controls the recordings ³. A vendor's help page describes its platform, not any legal right of access you may have where you live.

• the question as it appeared, word for word, and the answer you gave
• the date, the platform name, the employer, and the role as it was posted
• any notice you were given about AI being used, and the rejection message with its timestamp

Then raise it, in that order.

The employer first, because on a platform like HireVue the question came out of the employer's configuration, and a recruiter may be able to find out who wrote it. An enforcement body second, if the answer is bad or absent.

Which body depends on where you live and what you are claiming. Where to report an AI hiring decision you think discriminated walks the four doors and the filing clock.

At least one state has already said its existing law reaches this. In January 2025 the New Jersey Division on Civil Rights issued guidance on algorithmic discrimination, stating that the state's Law Against Discrimination applies to automated decision-making tools in employment the same way it applies to human conduct ⁶. A covered entity can violate it with no intent to discriminate, the guidance says, and a third party having built the tool is no defense ⁶. It closes by naming the state's own complaint route ⁶.

That is agency guidance interpreting a statute already on the books, not a new law, and it covers New Jersey only. It also does not turn the use of AI into a claim by itself: the complaint route still needs discrimination on a protected basis. What your own state requires before an interview is set out in your rights before an AI interview, by state and date.

Sometimes the answer is a question set written once and never reread.

Read the evidence >