The candidates walking away aren't rejecting AI

Thirty-eight percent of candidates have abandoned a hiring process after finding an AI interview in it. That number comes from a survey of 2,950 active job seekers published in May, and it usually gets read as a verdict on AI.

It isn't one. The same survey asked those people what they wanted changed. Nineteen percent said less AI. The rest wanted it kept or expanded, with conditions attached. The conditions are specific, and the largest of them is disclosure — being told, before anything starts, what is about to happen and who decides.

What they actually objected to

Among candidates who withdrew, 33% named a pre-recorded video interview scored by AI with no human present. 27% named not being told AI was involved. 26% named being monitored during the session.

Read those as process descriptions rather than as complaints. Not one of them is “a machine was involved in my application.” One is a format with nobody at the end of it. One is being told after the fact. One is surveillance with no stated purpose. Three decisions somebody made, and three that could have been made differently.

The disclosure failure is the widest of the three. 70% of candidates said they were never clearly told upfront that AI would evaluate them, and 21% found out only when the interview began. Just 18% said their employer had a clear AI policy at all. In the UK cut of the same data, 82% were not forewarned.

You can only disclose what your process actually does

Most teams treat disclosure as a writing task. Someone in talent ops is asked to add a line to the invite email, writes something reassuring, and sends it.

Try to write it properly and you find out it isn't a writing task.

Start with the sentence you need first: an AI conducts the first interview and scores it. Straightforward. Now the second: a member of our team reviews every score before any decision is made. Can you send that one? If a below-threshold score routes someone into a rejection template without a person ever opening the transcript, you can't. You would be putting a false statement in writing, to a candidate, in a jurisdiction that may later take an interest.

So it gets softened. Results are reviewed as part of our process. That is the sentence that produces distrust, because everyone recognises the shape of a sentence written to avoid saying who does what.

The third is harder: here is what we measure. Most teams can't finish it — not because the vendor won't say, but because nobody internally has asked. If you can't name the dimensions being scored, then the definition of a good engineer for your team now lives inside software that nobody on your side has read. That is a live risk in a market where first-round pass rates have been climbing for reasons that have nothing to do with candidates getting better.

The fourth carries legal weight: here is what we record and how long we keep it. India's DPDP Act governs consent for personal data, and its transitional compliance period ends in November 2026. Illinois requires that applicants be told AI will analyse the interview, be given an explanation of what general characteristics it evaluates, and consent before it runs — and that videos be deleted within 30 days of a request, backup copies included.

One more from that same statute, worth knowing before you design the flow: an employer relying solely on AI to decide who gets an in-person interview must collect applicant race and ethnicity data and report it to the state annually. Keeping a person in the decision is the defensible choice. In at least one jurisdiction it is also the cheaper one.

Every one of those sentences describes architecture, not tone. The notice is hard to write because writing it forces you to look at what you built.

The block I would send

Five statements, in the invitation, before the candidate schedules anything. Not on a policy page nobody will find.

• This interview is conducted by an AI interviewer. It asks follow-up questions based on your answers, so no two interviews run identically.
• It scores against [name the dimensions]. They are the same for every candidate applying to this role at this level.
• A member of our hiring team reads the transcript and the scores before any decision is made about your application.
• We record [audio / audio and video / your work in the exercise]. We keep it for [N] days, we use it only for this application, and you can ask us to delete it.
• If you would rather interview with a person, reply to this email and we will arrange it.

The bracketed parts are the whole exercise. Any one you cannot fill in is a finding, and the finding is worth more than the notice.

The last line is the one people resist, so be precise about what it costs. 46% of candidates said the option to request a human interviewer would make them more comfortable. In practice few request it, and the ones who do tend to be candidates with an accessibility need or a bad prior experience — the two groups you otherwise lose without ever seeing them. Offering the option costs you a handful of scheduled calls. Not offering it costs you people who close the tab, and they never show up in your funnel data because they never entered the funnel.

And if someone on your team is going to read those transcripts, it is worth being deliberate about what they are reading for.

The number that should bother you more than the 38%

Of the candidates who completed an AI interview, 28% advanced and 13% were told no. The other 51% heard nothing at all.

Silence after a screen predates all of this and is not an AI problem. But it is now filed under AI in the candidate's memory, and it is doing much of the reputational damage the technology gets blamed for. Someone who is told no, promptly, by a named person does not go and write a post about being processed by a machine.

34% said a poor AI interview experience left them with a worse view of the employer. 38% said a good one improved it. Most of what separates those two groups is whether anybody got back to them.

Write the notice first. Not as a compliance artifact — as a design review. The sentences you cannot finish are a list of the decisions in your hiring process that nobody made on purpose.