The Lookup Holds Until the AI Asks Its Follow-Up

What Does a Live AI Interviewer Do With Your Answer?

It builds the next question out of it. Alex, the AI recruiter formerly called Apriora, publishes candidate testimonials describing the AI summarising an answer back and then asking follow-ups based on that answer, with a human recruiter reading the result afterwards ³. That is marketing assembled from selected testimonials, so take it as a description of the design rather than evidence the design works.

Mercor puts a scale on the format. Its March 2026 engineering post says its interviewer, Monty, runs around 10,000 conversations a day, each fifteen minutes long ⁴. The interview is personalised before it starts by processing the candidate's resume, and coding problems are generated from the profile and the live conversation ⁴. Those volumes are self-reported, and Mercor's marketplace vets contractors for AI-training work rather than filling an employer's own roles ⁴.

The format is the part that carries over.

A recorded one-way interview asks a fixed list and waits.

A live one listens for a stopping point and then asks about the thing you just said, which is a different preparation problem. Four kinds of AI interview, and what each one records sorts out which one you were probably sent.

Why the Search Window Costs You the Answer

Because in at least one vendor's format, silence can end your turn before a result loads. Alex's employer-facing settings documentation carries an example welcome message telling candidates that after 2 seconds of silence their response automatically ends ¹. Employers can edit that text, so the exact figure will not hold everywhere, and the underlying cutoff is not documented separately ¹.

The same page describes an optional add-on the vendor calls Cheat Detection. It says the feature monitors for signals including background voices, excessive tab switching and unusual delays before answering, with the employer setting a flag threshold between 0 and 100 ¹.

It is an add-on, so plenty of interviews will not have it running. The page publishes no accuracy figure, no false-positive rate and no candidate-facing notice ¹.

Both of the confident answers here are wrong.

Nobody but the employer can tell you whether an interview is watching your tabs, and the employer is the one to ask.

What is more predictable is the cost inside the conversation. A lookup is likely to take longer than the pause you have, and what comes back is the answer a search engine would give anyone who typed the same words.

Employers read the same mechanism from the other chair. What employers are told about candidates using AI in live interviews describes assistance as strongest on questions with one public answer, and empty on a decision the candidate personally made.

Who Reads the Transcript Afterwards?

A person on the employer's hiring team, by one vendor's account. Alex's candidate FAQ says it does not make final hiring decisions and does not automatically reject candidates, and that its recordings, transcripts and notes are shared with the employer's hiring team, where humans make the decisions ². Alex wrote that description of Alex, and the account stops at the handover. What a hiring team then does with the transcript is not something the page can say.

So what gets read includes the conversation itself. A strong general answer sitting next to a thin specific one shows up on a page in a way it does not while you are talking. Employers do say they are looking for undisclosed use. In the Institute of Student Employers' Student Recruitment Survey 2025, using AI during interviews without disclosure or permission was the most commonly reported form of candidate misconduct, named by 61% of employers ⁵.

That records suspicion, not detection.

It covers UK graduate employers inside a self-selected membership body, and no verification step sits behind the figure ⁵.

Picture the two answers next to each other on a page. The first is fluent, unattributed and true of any employer, the sort of paragraph a search returns for the question as it was typed. The second, the one explaining why you chose that over the obvious alternative, comes out short and general, because there was nothing behind it to draw on. Neither answer proves anything by itself, and a reader does not need proof to prefer the candidate whose second answer had detail in it.

The ordering underneath is better evidenced than the catching. In a vignette experiment with 195 US panel members, a professional was trusted least when a third party exposed his AI use, more when he disclosed it himself, and most when nothing was said at all ⁶. The scenario was a tax adviser and his client rather than a candidate and an employer, so the ranking transfers and the sizes do not ⁶. The silent condition is an audience that never learned of any AI use, so it does not measure what hiding it pays.

Asking a recruiter what is allowed, before the round, is a short question and a fair one. The employer-side page on what follow-up questions expose about an answer is the plainest description of the check, written for the person running it.

Prepare Episodes That Survive Three Rounds of Why

Pick four or five things you actually did and take each one down three levels. Past-behaviour questions have a measured case behind them: Taylor and Small's meta-analysis of human-rated interviews found that past-behaviour questions with descriptively anchored rating scales produced a higher mean validity estimate than situational questions with anchored scales, .63 against .47 ⁷.

Those estimates rest on range-restriction assumptions that later work has called too generous, so the comparison between formats is the durable part and the absolute values are stale ⁷.

For each episode, write down four things: the decision you made, the option you ruled out, what made you rule it out, and what you would check before doing it again.

Take a migration you ran. Level one is what you moved and why anyone wanted it moved. Level two is the option you passed over and the constraint that killed it, a maintenance window nobody would give you, a library with no rollback path, a team that needed the old endpoints until the quarter closed. Level three is what you would verify first if you ran it again, and how you would know the verification had passed.

That is the level a probe reaches.

A search result stops one level above it, which is why the first answer holds and the second one comes apart.

Prep tools are useful here as long as they stay in the prep. What an interview copilot costs you on the follow-up makes the same case about live overlays, and the fix is identical in both places: do the work the night before, out loud, against the posting's own words.

See how it works >