Talk to an AI Interviewer in Finished Thoughts, Not Monologues

Why Does It Cut In When You Pause?

It is taking turns, and a pause is how it decides your turn ended. One vendor's employer documentation carries an example welcome message: when the interviewer stops speaking it is your turn, and after two seconds of silence your response ends automatically ¹. Employers can edit that text, so the exact number may differ from one deployment to the next, and the page does not document the underlying cutoff separately.

So a thinking pause and a finished answer look identical from the other side.

That may explain a good share of what feels like rudeness in a live round. You reach for the right word, the room goes quiet for a beat, and the next question lands on top of the sentence you had not finished. The same documentation describes an add-on called Cheat Detection, which an employer can switch on and which flags interviews on signals including background voices, excessive tab switching and unusual delays before answering ¹. Many interviews will not have it enabled, and the page publishes no accuracy information.

Which format you are in decides how much of this applies. A recorded one-way interview has nobody on the call to interrupt you and a fixed timer instead. Read what actually happens in an AI interview before you open the link, because the preparation is different for each kind.

Say the Shape of Your Answer Before You Give It

Announce the structure before a long answer, so the plan is on the record even if a pause ends your turn early. "There are two parts to this" tells a listener more is coming; a system that ends turns on silence may not wait for part two, but the transcript shows it was planned. Give the short answer first and the situation after. The habit costs nothing and it survives whichever kind of interviewer you end up with.

That is legibility rather than polish, and it helps a person reading the transcript too.

• End the sentence you are on. Trailing off into "so, yeah" is silence, and silence is the cutoff.
• Do the thinking before a sentence starts, not in the middle of one.
• When it cuts in, stop for a beat and let it finish the question.
• Keep a turn near a minute, then hand it back rather than filling the gap.

If it cuts you off mid-answer, say so and keep going. "I had not finished that one, let me add what happened" is an ordinary sentence in a conversation, and the transcript keeps it. Recovering out loud beats restarting from the top, because the second telling eats the time you had for the next question. Talking over it while it speaks is the move that reliably makes things worse, since both voices land in the same recording.

A different failure looks the same from the inside. When it asks something you already answered, or replies to words you did not say, the problem is the transcription rather than the timing. The AI interviewer misheard me and asked the same question twice covers that one.

Slowing down will not fix a microphone.

What Are the Follow-Ups Actually For?

They are the format rather than a trap, and several vendors describe them as part of the design. Alex's candidate page describes the AI asking role questions, summarising answers back, asking follow-ups from those responses, and letting candidates ask questions in return ². Mercor says its interviewer, which vets contractors for its own AI-training marketplace, is personalised from the resume before it starts, with coding problems generated from the profile and the live conversation ³.

HackerRank describes its AI interviewer offering real-time hints when a candidate is stuck and asking follow-up questions to judge thought process ⁴. Whether a given employer switched any of that on is its own decision.

All three are vendors describing their own products, and none of the three descriptions is an audit.

Employers read their own version of this. Whether candidates are using AI to answer questions during live interviews is the piece written for the other chair, and its argument is that a borrowed answer survives the first question and falls apart on the second. A follow-up does that work whether or not anyone borrowed anything.

That changes what preparation buys you. A rehearsed three-minute story has one shape and no branches, and a live round may stop it partway through to ask about the part you skipped. Prepared material still helps, but it has to survive being taken apart and put back in a different order. Answer the narrower question it actually asked instead of restarting the story.

What a follow-up question is looking for goes through the gaps a second question usually names, one at a time.

Ask your own questions when it offers. One vendor's page describes candidates doing exactly that inside a live AI round ². Ask what you would ask a recruiter: what the next stage is, who reads the recording, when to expect an answer.

When Is It Glitching Rather Than Following Up?

When the interruptions have no relationship to what you said. A follow-up quotes you back or narrows the question. A fault repeats itself, cuts out, or talks over silence. In a large randomised field experiment on AI voice interviews, reported in a working paper that has not been peer reviewed, the agent hit technical difficulties in 7 percent of its AI interviews ⁶.

That experiment randomised 70,000 applicants at one employer between a human recruiter and an AI voice agent, and 5 percent of applicants ended the call rather than speak to an AI ⁶. It was one firm in the Philippines running short customer-service interviews, so neither rate is a general failure rate for the format.

If it is genuinely broken, say so on the recording.

Name what happened in a sentence, out loud: the audio dropped for about ten seconds, and you are going to repeat that answer. On a platform like Alex, the recording and the transcript are what go to the hiring team, so a note left inside them can reach whoever reads it later. Then tell the recruiter afterwards, plainly and without a complaint attached. A short message describing the fault and offering to redo the round is a reasonable thing to send.

Alex's candidate FAQ says it does not make final hiring decisions and does not automatically reject candidates ⁵. The recording, transcript and notes go to the employer's hiring team, where humans make the decisions ⁵. That is a vendor describing its own product, and what the employer does with the summary sits outside the vendor's control.

In the 70,000-applicant experiment, human recruiters evaluated the transcripts and made every hiring decision in both arms ⁶.

The person you are actually talking to reads the transcript later, which is the reason to leave an honest note inside it.

See how it works >