How to detect AI-generated interview answers without accusing anyone

You cannot reliably tell an AI-generated interview answer from the words alone, and no tool can prove it: OpenAI withdrew its own AI-text classifier in July 2023 because of its low accuracy. What works is changing the question rather than analyzing the answer. Ask for details only someone who did the work would know, follow up two or three times, and notice whether the answer gets more specific or more generic. Treat anything you notice as a reason to probe, never as proof, and score only what the candidate actually showed.

This page is about answers that may have been prepared or generated with AI. A candidate reading from a script during a video call is covered in how to tell if a candidate is reading answers, and someone who is not who they say they are is a different problem, covered in how to spot candidate identity fraud.

Why can't you simply detect AI-written answers?

The detectors are not accurate enough. When OpenAI released its AI-text classifier in January 2023, it reported that, on a "challenge set" of English texts, the tool correctly identified 26% of AI-written text and wrongly labeled human-written text as AI-written 9% of the time. A note added to the same page says the classifier has not been available since July 20, 2023, because of its low rate of accuracy.

They are least accurate on the people you most need to be fair to. A study by Liang and colleagues published in Patterns (Cell Press, 2023) ran seven widely used detectors on 91 essays written by non-native English speakers for the TOEFL test and 88 essays by US eighth-graders. The detectors classified the US essays accurately but labeled more than half of the TOEFL essays as AI-generated, with an average false-positive rate of 61.3%. Plainer vocabulary reads as "machine-like" to these tools, which means a careful second-language speaker can look more suspicious than a fluent one using AI.

A live interview is not a text. A candidate can read an AI-written answer aloud, or use an app that listens to the call and shows answers in real time. As of October 2026, Cluely's site describes an assistant that listens in real time, gives instant answers and is invisible to screen share, so asking a candidate to share their screen proves very little.

What can make an answer sound AI-generated, and what else explains it?

None of the patterns below means a candidate used AI. Each has ordinary explanations, and the third column is the only part that should change what you do.

Which follow-up questions show whether the experience is real?

Real experience gets richer under questioning, including awkward details nobody would invent. Borrowed or generated material gets vaguer, repeats itself, or retreats to principles. Pick two or three of these after any answer that matters:

Specifics:      "Which system were you moving from, and why that one?"
Sequence:       "What did you do first, before anyone asked you to?"
People:         "Who disagreed with you, and what did they say?"
Numbers:        "Roughly how many customers did that affect?"
Mistakes:       "What went wrong in the first week?"
Trade-offs:     "What did you decide not to do, and do you regret it?"
Counterfactual: "If the budget had been half that, what changes?"
Teach-back:     "Explain it the way you'd explain it to a new hire."
Callback:       "Earlier you mentioned the migration. How did that
                 affect this project?"

The callback question is especially useful. A real career is connected; details from one answer show up in another, and a candidate who lived it can link them without effort. Pre-written answers rarely refer to each other.

How can you design an interview that rewards real experience?

• Ask about their resume, not the topic. "Tell me about the reporting rebuild you list at your last job" is harder to outsource than "How do you approach reporting?"
• Let the candidate choose the example, then probe it. A generic answer cannot anticipate which example they will pick.
• Use your own scenario. Describe a short, real situation from your team, with its constraints, and ask what they would check first. Context the candidate has just heard is hard to prepare for.
• Interrupt kindly. "Sorry, quick question before you go on: who signed that off?" A person telling their own story picks the thread back up easily.
• Score the follow-ups, not the opening. Write anchors that reward specifics and ownership, so a polished first answer with nothing behind it does not score well.
• State the rules in advance. Tell candidates what help is fine and what is not, so nobody is judged against an unwritten rule.

A sentence you can adapt for interview instructions:

You're welcome to prepare however you like, including with AI tools,
and to have your resume and a few notes in front of you. During the
interview itself, please answer in your own words without live help
from other people or AI tools. If you need an adjustment to how the
interview runs, tell us beforehand and we'll arrange it.

The run-of-show for remote interviews, including camera and connection checks, is in how to conduct a video interview.

What should you say if you suspect an answer is AI-generated?

Nothing that sounds like an accusation. You are not in a position to know, and a wrong accusation will end the interview and, for an agency, possibly the client relationship. Neutral lines that keep the interview going:

• "That's a good overview. Can you take me into one specific time it happened?"
• "Let's slow down on that one. What was your part, as opposed to the team's?"
• "It's fine if it's less polished. I'm more interested in what actually happened than the ideal answer."
• "I'd like to come back to that. Can you tell me about a time it didn't work?"

If the follow-ups produce real detail, you have your answer and the suspicion was wrong. If they do not, the candidate has not shown the capability, and that is what you score.

How do you record and score it fairly?

• Write what happened, not what you suspect. "Could not name the system or anyone involved after two follow-ups" is a fact. "Probably used ChatGPT" is a guess that does not belong in a candidate record; see what not to write in interview notes.
• Score against the anchors. An answer without specifics scores low whatever the reason, and an answer with rich, consistent detail scores high even if the opening sounded rehearsed.
• Do not adjust a score for suspicion. If you remain unsure, add a step for everyone at that stage, such as a work sample or a second interview on camera, rather than penalizing one person.
• Treat tool flags as prompts. Some interview tools flag answers that read as AI-written. Interview Signal's private answer check, for example, flags answers that read as AI-written or read out and suggests a question to try; it never changes a score and never goes into a client submittal. Any flag like that is a reason to ask a follow-up, nothing more.

The broader method of keeping evidence separate from judgment is in evidence-based interview feedback.

When is it more than AI help?

Using AI to prepare is common and usually fine. Two situations are different. A candidate who invents experience, with or without AI, is lying about their history, and how to handle a candidate who lies in an interview covers verification and next steps. A different person interviewing, or a manipulated face or voice, is identity fraud, which needs the checks in the identity fraud guide rather than better questions.

What about candidates who need adjustments?

Some candidates use notes, screen readers, captions, extra processing time or other assistive technology because of a disability, and some of those adjustments can look like the patterns in the table above. Ask every candidate in advance whether they need an adjustment, agree it before the interview, and tell the interviewers so they do not misread it. ADA accommodations in interviews explains the process.