Are Candidates Using AI During Live Interviews?

Which tools are candidates actually using?

The category is called an interview copilot, and the products are real businesses. Cluely, built by the two Columbia students behind Interview Coder, raised a $15 million Series A led by Andreessen Horowitz in June 2025, at a valuation investors put near $120 million ¹. What it sells is a hidden in-browser window that the interviewer or test giver cannot view ⁷.

That mechanic decides what you can do about it. A window the call cannot see is also a window screen sharing does not show you, because screen sharing shows the surface the candidate chose to share. The rest follows from where the software runs rather than from anything a vendor has published: what sits on the candidate's own display never crosses to your side of the call, so there is nothing in your video vendor's logs to find.

Underneath the named products sits a much larger grey market that needs no product at all: a second laptop, a phone propped below the camera, a chat window open behind the call. Above them sits identity fraud proper. The FBI's advisory on deepfaked applicants for remote positions describes interviews where lip movement does not coordinate with the audio, and where coughing or sneezing does not line up with what is on screen ². That is a different problem with different remedies, and conflating the two is how interviewers end up policing nerves.

So the honest answer to whether it is happening is yes, at a rate you cannot measure, in a round that gives you nothing to measure it with. Which makes the useful question not how many, but which of your questions are answerable this way.

Which questions can an assistant answer for them?

Anything with a canonical answer. A definition, a complexity bound, the difference between two named patterns, the standard framework for a market-entry case, the textbook tradeoff between two architectures. Each has a public right answer, and a model returns it in seconds, phrased better than most candidates manage under pressure. A definitions-heavy screen for an engineer or an analyst is now close to uninformative.

Sort your current question set into two piles. The first pile is everything a well-read stranger could answer without knowing the candidate:

• What is the difference between X and Y?
• How would you approach a problem of this shape?
• Which metric would you track for a product like this?
• Walk me through how that algorithm, framework or process works.
• What are the tradeoffs of this design?

That pile was already the weak half before any of this. It rewarded preparation over practice, and interview-prep sites published its answers a decade ago. What changed is that the preparation no longer has to happen in advance, which removes the last thing the question was accidentally measuring, which was whether the candidate cared enough to study. If your loop leans on this pile, the polish is not a candidate signal, and uniformly excellent answers are the tell.

The same holds for a live coding round built on a problem with a published solution. The realistic response is not to hunt for a harder puzzle, which lasts one quarter, but to open the assistant deliberately and watch what the candidate does with it. Running a coding interview with AI open is its own piece of work.

Which questions are unhelpable?

Questions about something the candidate personally did. A decision they made, what they knew when they made it, what they traded away, what happened afterward, and what they would do differently. No assistant holds that record. OPM defines a structured interview as systematically asking about a candidate's behavior in past experiences, their proposed behavior in hypothetical situations, or both ³, and the past-experience half is the half that still holds.

Four that travel across roles, because each needs a fact only the candidate has:

• The undo. Tell me about something you shipped and then had to pull back. What made you pull it?
• The number you stopped trusting. When did you last decide a figure in your own analysis was wrong, and how did you settle it?
• The thing you should have kept. What did you hand to a model in the last month that you should have done yourself, and how did you find out?
• The disagreement. Name a call your manager made that you thought was wrong. What did you do about it?

An assistant will generate a plausible story for any of these. What it cannot do is survive the second question, because the second question depends on what the candidate just said. Ask which number moved, by how much, and who else saw it. A retrieved answer has depth of exactly one, and the pause before the probe is the part that shows. Follow-up questions expose understanding faster than any transcript analysis will.

One more family is worth adding: questions about work the candidate has already submitted. Hand back their take-home or a portfolio piece, change one constraint, and ask what breaks. That cannot be prepared and cannot be retrieved, and anyone who did the work answers it in ninety seconds.

Should you watch for tells instead?

No, not as the main instrument. The cues people trust run backwards. Across six experiments with 4,600 participants, people identified AI-generated professional self-presentations at 50-52% accuracy, barely above chance, and the texts carrying grammatical problems (which participants were more likely to rate as machine-written) were in fact 15% less likely to be AI-generated ⁴.

The live-interview tells are worse, because each has an innocent cause more common than the guilty one. Eyes moving off camera: notes, a second monitor, a nervous habit, an attention difference. A pause before answering: processing in a second language, a bad connection, a stutter, thinking. A register that turns suddenly formal: a rehearsed answer, which is the thing you asked them to prepare. Eyes off camera are not evidence on their own.

Running the transcript through a detector afterwards does not rescue it. An independent review of fourteen detection tools found them neither accurate nor reliable, biased toward calling text human-written, and badly degraded by ordinary rewording ⁶. On a speech transcript the input is already a machine's guess at what was said. The longer version is in whether AI detectors work on interview transcripts.

Exposure is the other cost. A suspicion you act on is a selection procedure, and EEOC guidance holds that a procedure used to make an employment decision must be job-related and consistent with business necessity where it screens out a protected group ⁵. "The answer sounded too smooth" is job-related to nothing, and the people it screens out are the ones who prepared hardest and the ones whose speech does not match your default.

Change the round, not the surveillance

Two changes carry most of the weight. State the rule in the invitation, in one sentence, so nobody has to guess. An unstated rule gets guessed at, and the guessing tracks how much coaching a candidate has had. Then rewrite the question set so most of it asks about work the candidate has already done, with the follow-up written before the round rather than improvised in it.

The rest is bookkeeping:

• Pick a side and publish it. Banned, allowed, or allowed and scored are all defensible; unstated is not. Letting candidates use AI in the interview is a decision to make once, not per interviewer.
• Score against a written key. Decide what a good answer contains before the first candidate, apply it to everyone, and record why each answer landed where it did. That record survives a challenge. A hunch does not.
• Give every question one prepared probe. The probe is the instrument; the question is only the setup.
• Move canonical-answer questions out of the live round. If a fact matters, check it in writing in two minutes, or accept that it gets checked on the job.

What none of this buys you is the thing the interview was standing in for: whether this person works well with an assistant. A live round can capture a candidate describing how they would check a confident claim. It cannot capture them checking one, because there is nothing in the room to check against. That gap is what a work sample with the assistant open is for. See how Olive measures this.

See how it works >