How the answer check works in Interview Signal

The answer check is a private flag on Interview Signal's live scorecard. When a candidate's answer reads as if it was generated by an AI tool or read out from one, rather than recalled, a card appears saying "This answer may be AI-assisted" or "This answer reads as AI-generated", with the reasons and a question to try. It never changes a score, it is never written into the submittal, and it stays on your computer. It is on by default and can be switched off. This describes version 0.9.10, current as of October 2026.

When does the answer check run?

It runs as part of scoring, in the same request, so it adds no extra step. Scoring happens when you press Evaluate, and automatically with Score as we go once the candidate has said about ninety new words. If you turn Score as we go off, the answer check only runs when you press Evaluate. With Dual AI on, both models give a reading and the stronger one is shown.

What does a flag look like?

A card appears above the scorecard with:

• A heading: "This answer may be AI-assisted" at the medium level, or "This answer reads as AI-generated" at the high level.
• When it happened, as the elapsed time (for example "at 14:05"), how many flags there have been so far, and "both models" when two models agreed.
• The reasons, as short phrases naming what in the words raised it.
• A try asking question: one that someone who really did the work could answer at once and a script could not, about a specific detail of what they just claimed.
• The footer "A reason to probe, not proof. It does not change the score." and an × to dismiss it.

Answers are read at four levels: none, low, medium and high. Only medium and high are shown. Low is recorded but not shown, because a flag that fires on every articulate answer trains people to ignore it.

What makes an answer get flagged?

The check judges only the words. Its signals count only when several appear together:

• a textbook or documentation-style definition where a personal account was asked for;
• a tidy framework or list ("firstly, secondly, in conclusion") with no event, name, number, date, tool version or mistake of their own in it;
• a sudden change of register from how they spoke earlier in the same interview;
• restating the question, then a generic answer that would fit any company;
• confident coverage of every angle, with nothing only someone who was there would know;
• claims that contradict something they said earlier.

A medium or high level also requires a sign of polish, such as a formal or written register, a recited structure or tidy coverage of every angle, and one of the reasons must name it.

What never raises a flag?

The check is instructed never to raise a level for an accent or non-native phrasing, nerves, a well-prepared answer to a common question, being articulate, speaking slowly, or short answers. Vague is not generated: a hedging or evasive answer such as "it depends" or "we did a lot" is a weak answer and scores low, but its answer-check level is none. Clichés in casual speech and giving the team the credit are treated the same way. A rehearsed true story is not a generated one.

What should I do when an answer is flagged?

1. Ask the try asking question, or your own question about a specific detail: who else was involved, what went wrong, what they would change.
2. Listen for first-hand detail. Someone who did the work can usually add it without pausing.
3. Score what they showed. The scorer already treats polished general claims as thin evidence, so a fluent answer with nothing checkable in it scores low whether or not it is flagged.
4. Do not accuse anyone on the call. Note what you observed and follow your agency's or company's policy.

Our guides to detecting AI-generated interview answers and telling if a candidate is reading answers cover fair follow-up technique, and what interview fraud is covers the wider problem.

Where do flags go after the call?

The review screen lists them under Answer checks, with the time, the level and the reasons, and this note: "For you only — not in the scorecard or the submittal. A reason to have probed, not proof." Flags are stored in the interview's folder on your computer. They are not in any export and not among the counts sent to Interview Signal; see where your data is stored.

How do I turn it off?

Untick Flag answers that may be AI-generated in Settings under "AI during interviews", or in the AI menu on the live screen during a call. The choice is remembered. While it is off, scoring requests do not ask for the check at all.

Limits

• It reads the transcript only. It cannot see the candidate, their screen or their eyes, and it does not analyze the audio.
• It can be wrong in both directions: a polished true answer can be flagged, and a well-disguised generated one can pass.
• Transcription errors feed into it, since it reads what was transcribed.
• It is not a detector to act on alone. Interview Signal is not legal advice; decide on documented answers, applied the same way to every candidate.