What to Do When a Candidate Admits to Using AI in an Interview

The moment that catches teams off-guard

Halfway through a take-home walk-through, the candidate says it: "I used ChatGPT for parts of this." Or in a live interview, after a paste shows up in the integrity log, they volunteer: "Yeah, I had Copilot on — I always do."

Most hiring teams have no written policy for this moment. The interviewer freezes, ends the round politely, and the panel argues in the debrief. Some teams reflexively disqualify. Others wave it through because "honesty counts." Both reactions are wrong because both skip the only question that matters: can this candidate do the work without the AI doing it for them?

This post is the playbook we recommend when a candidate volunteers an AI admission, drawn from how teams using integrity-first hiring actually handle the conversation.

Step 1: Don't react. Acknowledge and continue.

The first 10 seconds set the tone. If you flinch, the candidate clamps up and you lose access to the only thing that will help you decide — their honest account of how they used it.

A neutral acknowledgement works:

"Thanks for telling me. AI use is something we think about a lot, and I'd rather understand exactly how you used it than guess. Walk me through it."

You are not granting permission retroactively. You are widening the conversation so you get real information. The candidate's relief usually produces a more accurate account than any forensic analysis you could run later.

Step 2: Find out which mode of AI use you're dealing with

There are three meaningfully different cases. Most disqualifications happen because teams collapse them into one.

• Mode A — AI as accelerator. They used Copilot inline for boilerplate, asked ChatGPT to remind them of a syntax detail, or had Cursor regenerate a function they then rewrote. They drove. The AI saved keystrokes.
• Mode B — AI as collaborator. They wrote the prompt, evaluated multiple outputs, picked one, debugged it, and integrated it. They can defend every architectural choice but not necessarily every line.
• Mode C — AI as ghostwriter. They typed the problem into the model, pasted the answer back, and only touched the code to make it run. They can describe what it does but cannot explain why it is shaped this way or what would change under different constraints.

Mode A is how most working engineers code in 2026. Mode C is the problem. Mode B is where judgment lives — and where the candidate's follow-up answers are doing most of the work in your decision.

Step 3: Probe with mode-discriminating questions

Asking "did you use AI?" is useless after the admission — you already know. The questions that actually separate the three modes:

1. "Show me where you pushed back on the AI's first suggestion." Mode B candidates can point to a specific moment. Mode C candidates either freeze or describe a generic "I changed variable names."
2. "What's the part of this you couldn't have written without the AI? What would you have done instead?" Mode A and B can answer cleanly. Mode C reveals itself through vague answers or a sudden defensive tone.
3. "Walk me through what happens when this function gets called with [a specific edge case from the code]." This is the highest-signal question in the entire interview. The candidate has to simulate execution. Mode A and B trace it; Mode C stumbles.
4. "If the constraint changed to [X], what would you change?" Tests whether they reasoned about the design or just accepted the AI's defaults. Mode C candidates often answer "I'd ask the AI."

A candidate who answers these crisply is showing you Mode A or B regardless of how they wrote the first draft. A candidate who can't is showing you Mode C regardless of how clean the artifact looks.

Step 4: Cross-reference with the integrity signals you already have

This is where the integrity report earns its keep. The admission gives you the candidate's narrative; the signals tell you whether the narrative matches the session.

For a live coding round, look at:

• Keystroke rhythm. Mode A produces bursty-but-human typing with thinking pauses. Mode C produces long silences punctuated by single large paste events.
• Paste-event timing. A paste right after a 30-second silence followed by minor edits is the classic Mode C signature. Pastes interleaved with active typing are usually Mode A.
• Code coherence. Does the writing style stay consistent across the file, or does it shift between functions in a way that suggests different authors? Inconsistency lines up with Mode C; consistency with A or B.
• Time-to-first-meaningful-keystroke. A candidate who started typing a working solution within 8 seconds of seeing a non-trivial problem was not thinking.

The point is not to "catch" the candidate. They already told you. The point is to map their admission to specific evidence so the panel debrief has facts, not vibes.

Step 5: Decide against the role's actual standard

Now you have three inputs: the admission, the follow-up answers, the signals. Match them to the role.

• For most engineering roles, a candidate in Mode A or B who can defend their submission is a hire-able candidate. The market reality is that the engineers you want are using AI daily. Penalizing them for being honest about it while hiring the candidate who used AI silently is the worst possible outcome.
• For roles where independent reasoning is the explicit bar (research engineer, security analyst, anything where the candidate has to think through unfamiliar problems with no assistant), Mode C is a no-hire regardless of artifact quality. The artifact is not what you are hiring.
• For senior roles, the test is whether they can articulate the trade-offs the AI made on their behalf. A senior engineer who used AI to scaffold a solution and can defend the architectural choices is fine. A senior engineer whose code is fluent but who cannot answer "why this approach over X" is showing you they are operating below the level you are paying for.

If you cannot decide, schedule a 30-minute follow-up with a different problem and an explicit "no AI on this one" policy. Most candidates pass cleanly. The ones who don't gave you your answer.

Step 6: Update your policy so this doesn't happen ad hoc

A candidate admission catches teams off-guard because the policy was unwritten. Fix that upstream. In every assessment invite, state:

• For async take-home rounds: "Use whatever tools help you. You'll do a 30-minute walk-through where we ask you to explain your code."
• For live coding rounds: "Please don't use AI coding assistants during this round. Docs, notes, and your normal IDE are fine."

When the rules are explicit, an admission either confirms a candidate followed them (good signal) or reveals they did not (real data, not an ambiguous moment).

What to do next

The candidate who admits to AI use is doing you a favor. They are giving you information that the candidate who used AI silently — and there are many — is hiding. Your job is to use the admission as a starting point for a sharper conversation, not as a shortcut to a verdict.

The teams that get this right share three habits: written AI policies per round, a walk-through that probes for reasoning depth, and an integrity layer that confirms or contradicts the candidate's account. With those in place, an AI admission becomes one signal among many — exactly as it should be.