When the AI Interviewer Loops, Say So and Keep Answering

What Do You Say in the Next Ten Seconds?

Say the plain thing: that the question did not come through, and that you are going to answer what you believe was asked. Then answer it. The recording is running, so that sentence is now part of the record a person reads afterwards, and it turns a confusing stretch of audio into something a reviewer can place.

On Alex, the AI interviewer formerly called Apriora, the vendor's candidate FAQ says the recording, the transcript and the notes all go to the employer's hiring team, and that the people there make the decisions ¹. That is a vendor describing its own product, which is not an audit of it. But it tells you who your sentence is for. You are not explaining yourself to the agent that just malfunctioned. You are leaving a note for whoever opens the file next week.

Ten seconds of clear narration beats a minute of guessing.

Keep it factual and short. "That came through as the same question you just asked, so I'll answer it as I heard it" is enough. There's no need to apologize for the software, and no need to fill the gap with commentary about the tool.

Then go back to the answer you were giving.

If it got a name wrong, correct it once and move on. A transcript with your last employer spelled three ways is easier for a reviewer to read past than an answer that never arrived.

How Often Does This Actually Happen?

Often enough to have been measured. In a field experiment at a Philippines customer-service employer, 70,000 applicants were randomly assigned to a human recruiter or an AI voice agent, and the agent hit technical difficulties in 7% of cases ². Human recruiters read the transcripts and made every hiring decision in both arms of that study.

That is one firm, one country and one job family: high-volume customer-service hiring, in interviews of ten to twenty minutes. It is a 2026 working paper, not a peer-reviewed result. So 7% is not a number to quote about your own interview, only enough to say a glitch is ordinary.

A loop is far more likely a bug than a stress test.

That matters for what you do in the next minute. If you read it as a test, the temptation is to over-explain, and the over-explaining is what eats the answer you were giving. Treat it like a dropped call instead: name it, carry on, and put the substance back where it was going.

A glitch is a known category of fault rather than a story about you. That is the frame worth keeping when you write to the recruiter later that day, and it is why the note works better as a fact than as a grievance.

Send the Recruiter a Short Factual Note

Write to the recruiter the same day, and keep it to what happened. Name the question, the approximate time, and what the software did, then ask what the process is for a technical fault. A message that reads like an incident report gives a coordinator something to forward. A message that reads like a complaint about the tool gives them nothing to do.

Write the time down while it is fresh. A timestamp is the most useful thing you own afterwards, because it lets somebody find the exact stretch of a long recording.

The closest published route is narrower than a redo. Alex's candidate FAQ says a candidate who misses a slot because of a technical issue can take the interview later using the same link ³. That covers a session you could not join rather than one that glitched partway through, and it is one vendor's policy rather than a rule every employer follows.

So ask by name for what you actually want, and say plainly that the interview ran and broke. The worst answer is no, and a no still tells you the first recording is the one on file.

Ask what the employer does when an interview breaks partway through.

Keep the note free of theories about why the system failed. You don't know, and the recruiter probably doesn't either. What they can do is check the recording, ask the vendor, or offer a second attempt, and all three start from the same two facts: which question, and when.

The employer-side piece on the neighbouring question, what an employer owes a candidate the AI screened out, covers correcting wrong data and a person who can look again, from the hiring team's chair.

What If the Glitch Changed the Outcome?

Ask for a human review, and ask to see the transcript. Several platforms publish something for the first of those: one routes an unscorable answer to a person on its own, and two let you ask. HireVue's explainability statement carves out the answers its AI does not score at all: one the system cannot make out clearly enough, or one hit by a microphone problem, goes to a person for review instead ⁴.

Landing in that queue is not the same as scoring well. It moves the response to a person and says nothing about what that person concludes, and the vendor publishes no figure for how often it happens. HackerRank's candidate AI notice says you may request human review of a decision made or substantially assisted by AI ⁵. It also lists an explanation of how AI was used in an adverse decision. At micro1, the candidate privacy notice gives an email address for appealing a decision based on automated technologies ⁶.

Each is a request, not a guarantee.

Both notices frame these as things you may ask for. HackerRank's points the request at the employer rather than at the platform. The micro1 notice points it at micro1's own privacy address. Neither says the request has to be granted. Where no law where you live requires it, the answer rests on the policy of whoever receives it, which is a reason to ask early and in writing rather than a reason to skip it.

If the transcript is what gets read, it is worth asking to see it. A 2024 audit of one version of OpenAI's Whisper found about 1% of transcription segments carrying whole phrases or sentences that appeared nowhere in the audio, and the invented lines fell hardest on speakers who paused longer ⁷. That audit used a research corpus of speech from people with aphasia, not job interviews, and the same paper found nothing comparable from four other commercial services on the same segments.

A platform may not say which engine it uses, and a fabricated line reads exactly like a real one on the page.

Two companion pieces cover the ask itself: does a human watch your AI interview and asking for a human review.

See how it works >