How to spot AI-assisted answers in a remote interview

Every recruiter running remote interviews in 2026 has had the same uneasy moment. The answer is fluent, well-structured, and covers every point you hoped for — and something about it feels borrowed. You cannot say what, and you certainly cannot say it out loud.

Real-time AI assistants that listen to an interview and feed a candidate answers are now cheap, plentiful, and specifically marketed for this. Pretending otherwise is not a strategy. But neither is paranoia: falsely suspecting a nervous, over-prepared candidate is a worse outcome than missing an assisted one, and it will cost you good hires.

What follows is what actually discriminates, and what only feels like it does.

The signals that do not work

Start here, because acting on these is how interviewers end up rejecting honest people.

• Looking away from the camera. People look away to think. They also look away because their second monitor has the job description on it, because they are in a shared flat, or because sustained eye contact over video is unnatural. This tells you nothing.
• Pausing before answering. A thoughtful candidate pauses. So does someone reading. The pause itself is not evidence — what matters is what the pause produces, which we will come back to.
• Speaking in polished, structured prose. Some people genuinely talk like that, particularly consultants, senior engineers who write a lot of design docs, and anyone who has rehearsed. Polish is a weak signal on its own.
• Using industry vocabulary correctly. Knowing the words is the baseline expectation for the role. Penalising fluency is how you end up hiring the least articulate candidate in the pool.

The signals that do work

Assisted answers fail in a specific and consistent way: they are generically correct but personally empty. The model knows the domain. It does not know the candidate's last four years.

1. Textbook completeness with no situational mess

Ask an experienced engineer how they handled a production incident and you get mess: the wrong hypothesis they chased for forty minutes, the colleague who spotted it, the fix that was ugly but shipped at 2am. Real experience is textured and slightly unflattering.

An assisted answer gives you the clean version — the five canonical steps of incident response, in order, with nothing embarrassing in it. When someone's history sounds like a best-practices document, that is worth probing.

2. Latency that scales with question novelty

This is the most reliable behavioural signal, and it is about the pattern, not any single pause. Track roughly how long the candidate takes to start answering across the whole conversation.

A candidate drawing on their own memory is fastest on questions about their own work and slowest on abstract or hypothetical ones. A candidate reading from an assistant shows the opposite shape — smooth on general knowledge questions, and unexpectedly slow on the ones about their own résumé, because the model has to be told what happened before it can help.

3. Answers that do not inherit the previous answer

Human conversation carries context forward. If a candidate told you ten minutes ago that their team was four people and had no dedicated QA, a later answer about their release process should reflect that. Assisted answers often reset — each response is locally excellent and globally inconsistent, because the assistant is answering a question rather than continuing a conversation.

Deliberately ask a question whose only good answer depends on something they said earlier. Inconsistency here is far more diagnostic than any amount of eye-tracking.

4. Depth that collapses exactly one layer down

An assistant will confidently explain what a technique is and why it is used. Where it thins out is the second and third follow-up into a specific decision the candidate personally made.

• "You said you moved to event-driven processing — what did you have to give up to get there?"
• "Who disagreed with that decision, and what was their argument?"
• "What would you do differently if you started that project again on Monday?"
• "What broke first after you shipped it?"

These have no textbook answer. Someone who lived the project answers them faster than the technical questions, often with visible feeling. Someone who did not goes abstract, hedges, or redirects to a general principle.

How to probe without accusing

You will never confirm AI assistance by confronting someone, and attempting it poisons the interview and your employer brand. The move is to make assistance useless rather than to catch it.

1. Anchor every question in their specific history. "How would you design a rate limiter" is trivially assistable. "Walk me through the rate limiting you built at your last company, and what you got wrong first" is not.
2. Interrupt politely and often. Assistants perform badly when the question changes mid-answer. "Sorry — before you go on, when you say the queue backed up, how far behind did it actually get?" is a normal conversational move and a very effective one.
3. Ask for numbers, then ask what the numbers meant. Fabricated metrics rarely survive a follow-up about why that number was the one worth watching.
4. Go sideways, not deeper. Instead of pushing further into a topic they are handling well, jump to an adjacent decision. "You've covered the migration — who was most affected by it, and how did you find out?"
5. Leave silence after a thin answer. Do not fill it. People who know more will add more; people who do not will repeat themselves in different words.

Decide in advance what you would do about it

This is the part most teams skip, and it is the part that matters. Suspicion without a policy turns into inconsistent, unfair, and probably discriminatory decisions.

Write down, before the interview, what your position actually is. There is a defensible range:

• Assistance is disqualifying, stated explicitly to candidates up front, applied to everyone.
• Assistance is irrelevant on knowledge questions but disqualifying on questions about the candidate's own history, because that is a truthfulness issue rather than a tooling one.
• Assistance is fine, and the interview is redesigned so it does not help — every question is about what this specific person did, decided, and regretted.

All three are legitimate. What is not legitimate is deciding case by case based on a feeling, after the fact. If you cannot articulate the rule to a candidate, it is not a rule — it is a bias.

Where tooling helps

Doing all of the above while also conducting a warm, human conversation is genuinely hard. You are supposed to track answer latency across twenty questions, remember what they said ten minutes ago, notice that a claim contradicts an earlier one, and think of the right sideways follow-up — in real time, while making the candidate feel comfortable.

That is the case for having something listen alongside you. Hiriso follows the conversation live, flags answers that pattern-match AI assistance, surfaces the inconsistency you would otherwise catch only on the replay, and suggests the follow-up that pins a claim down — while you stay focused on the person.

It does not make the decision, and it should not. It gives you the second pair of ears that makes the judgment yours to make properly.