Can candidates cheat an AI interview?

Yes. Some candidates already do, and the tools they use are cheap and hard to spot.

We built an AI interviewer and decided not to build a cheating detector. This post explains why, and what we do instead.

Yes, and detection is not the fix

A candidate can run an AI copilot during almost any interview format. It listens to the question, drafts an answer and shows it on a hidden overlay or second phone.

Vendors claim the best ones display that text in about a second. So the honest answer to the question in the title is yes.

The more useful question is what a hiring team should do about it. Our answer is not “detect it”. The rest of this post argues that case.

How candidates actually use AI

The tools have names now. Cluely launched in 2025 with the slogan “cheat on everything”. Later that year it raised a $15 million Series A from a16z.

Final Round AI, Verve AI, Parakeet and LockedIn AI sell live interview assistance.

The mechanics are the same across all of them. Capture the audio and transcribe it live. Send it to a language model. Show the answer where the interviewer cannot see.

The vendors compete on invisibility. Verve AI markets a stealth mode built to stay hidden during screen sharing.

Tools such as Cluely use overlays that screen sharing may not capture, according to Fabric’s explanation published in 2026.

Several ship a mode for phone interviews. The copilot runs on a second device and the candidate reads from it.

How many people do this? Resume Genius surveyed 1,000 US job seekers. On 4 May 2026 it reported that 22% had used AI in a live interview.

That is self-reported, so treat it as one survey result rather than a prevalence estimate.

Are AI interview assistants detectable?

Sometimes. A flag needs supporting evidence and human review.

The copilot vendors and the detection vendors are in an arms race. Overlays learned to dodge screen capture, so proctoring moved to the operating-system layer.

Then the copilots moved to a second device, which no software on the candidate’s laptop can see.

What is left are behavioural tells. A flat pause before every answer. Eyes tracking left to right. Answers with perfect structure and no specifics. Trouble with follow-ups.

Every one of those has an innocent explanation. Nerves, a second language, neurodivergence or well-rehearsed notes produce the same signs.

Karat has run more than 500,000 technical interviews. In March 2026 it put it plainly: there is rarely a single definitive sign.

What detection catches and what it misses

Fabric, a company that sells cheating detection, published its own numbers in January 2026. Across 19,368 interviews on its platform, 38.5% were flagged for suspected AI use.

That is a flag rate on one vendor’s platform, not a measurement of how many people cheat.

The same report says 61% of flagged candidates still scored above its pass bar. It does not say whether any hiring decision changed because of a flag.

A signal that fires on a majority of people who then pass is hard to act on.

HireVue tells a similar story from the other direction. It announced in January 2021 that it had removed visual analysis from new assessments in early 2020; a complaint to the US Federal Trade Commission had been filed in 2019.

HireVue says its remaining integrity signals go to a human reviewer, never to an automatic rejection. Even a detection vendor will not let a flag make the decision.

The false positive nobody prices

Detection vendors talk about catch rates. The people wrongly caught deserve the same attention.

The best-documented case is text. OpenAI launched a classifier for AI-written text in January 2023.

It identified 26% of AI text correctly and mislabelled 9% of human text as AI. OpenAI withdrew it six months later “due to its low rate of accuracy”.

Worse, the errors were not random. In 2023, Liang and colleagues tested seven GPT detectors on essays by non-native English writers.

The detectors flagged most of them as AI-generated while largely passing the native writers in the comparison group. A prompt asking for more elaborate vocabulary reduced the false positives rather than removing them; a separate self-edit prompt helped AI-written essays evade detection.

That study tested text, not live interviews. Its results do not establish the error rates of live-interview detectors.

Assume a flag trained on “typical” answers will misfire on atypical people.

Ask any vendor for their false-positive rate by language group. If they cannot give you one, you are buying bias risk and calling it integrity.

What the law lets you watch

Even if detection worked, much of it is now off the table in the EU, and increasingly in the US.

The EU AI Act bans AI emotion inference based on biometric data in the workplace, including recruitment, with exceptions for medical or safety uses. Article 5(1)(f) has applied since 2 February 2025.

The maximum fine is 35 million euros or 7% of global turnover. Some detection products still advertise stress or sentiment analysis. Whether those features fall under this ban depends on how they work; text-only sentiment analysis is not covered by this specific prohibition.

GDPR adds a proportionality test to everything else. In 2021, Italy’s data protection authority fined Bocconi University 200,000 euros over exam proctoring.

The findings included unlawful biometric processing, profiling and consent that was not freely given. Access conditioned on monitoring may not produce valid consent, and an interview is access too.

The United States is a patchwork, but the direction is the same. For jobs based in Illinois, employers using AI to analyse recorded video interviews must provide notice and an explanation and obtain consent before asking applicants to record them.

New York City requires an annual independent bias audit of covered automated employment decision tools.

Our reading, not legal advice: assessing what a candidate says needs a lawful basis and clear notice. It also needs a data-protection review of your own case.

Inferring how they feel, or watching their eyes and their screen, needs a justification you could defend to a regulator.

Preparing with AI is not cheating

The “cheating” framing blurs a distinction worth keeping.

The same self-reported Resume Genius survey said 78% of job seekers had used AI in the application process or would consider using it. Common uses were rewriting a CV, rehearsing answers and researching the company.

That is preparation. Candidates have always done it with whatever tools existed.

Feeding answers live is different. That is the narrow behaviour worth designing against, and the one your policy should name.

Most policies do not. Greenhouse’s April 2026 report covered 2,950 job seekers across five countries; among US candidates with relevant experience, 80% said employer AI policies were vague, rare or absent.

Publish the line before the interview. Prepare however you like. Answer in your own words.

Design questions a copilot answers badly

If detection is weak and monitoring is legally boxed in, the lever left is the interview itself. Fortunately it is the strongest lever there was.

Structured interviews are among the best-validated predictors of job performance. Sackett, Zhang, Berry and Lievens re-analysed the selection literature in 2022 and discussed the findings in a 2023 follow-up.

They put the mean validity at .42, first among widely used predictors, ahead of cognitive ability tests. Structure is what moves the number.

Structure also happens to be hard to fake. A few design choices do most of the work:

• Ask about the candidate’s own experience, then follow up on the detail. Follow-ups can test specific details, but a copilot can also invent them.
• Use scenarios specific to your job. If the answer is in a model’s training data, the question was too generic to begin with.
• Score against written criteria, not against how polished the answer sounded. Polish is exactly what a copilot supplies.
• Treat screening as a filter, not a verdict. The next stage, a work sample or a conversation with the hiring manager, closes the loop.

Why Kira has no proctoring

Kira is a voice interview that runs in the browser. There is no video, so there is nothing to read from a face and no gaze to track.

There is no screen sharing, so there is nothing to monitor. Kira holds the audio and a transcript with silence markers. It does not record video or your screen.

We considered building detection and decided against it, for the reasons above. The catch rate is unproven and the false positives fall on the wrong people.

In EU hiring, AI emotion inference from biometric data is generally banned, with exceptions for medical or safety uses. The rest of the monitoring needs a defensible basis we did not want to construct.

Voice interviews do not prevent the use of an AI copilot. A candidate can still read suggested answers from an overlay or second device.

Reading an AI-generated answer aloud does not reliably reveal that AI was used.

When an answer to an interview question is vague, Kira asks one short follow-up for a specific example. She never follows up on must-have checks.

A scripted answer tends to run out at the follow-up.

We do not claim this catches anyone. It is a design that makes the shortcut less useful, which is a different thing from surveillance.

What we do instead

Everything Kira produces is built to be checked by a person.

Every quote on a scorecard is matched against the real transcript before it is shown. If Kira cannot verify a quote, it does not appear.

When the evidence for a criterion is thin, Kira leaves that criterion unscored rather than inventing a score. The scorecard shows that state openly.

Kira never makes the hiring decision. The AI assessment is read-only. Your team records its decision separately.

Kira never rejects a candidate. A recruiter reads the transcript, the timestamped quotes and the per-criterion anchors, and makes the call.

We chose visible evidence and human review over cheating detection. It is the less exciting option, and the one we can defend.

FAQ

Can candidates use ChatGPT during an AI interview?

Yes. Copilot apps transcribe the question and draft an answer, in about a second by vendors’ own best-case claims. Screen-based formats give them an overlay to hide behind. That includes browser-based voice interviews: candidates can still read suggestions from an overlay or second device.

Does Kira detect cheating?

No. Kira has no proctoring, no video and no screen monitoring. It records and transcribes what the candidate says. Every scorecard quote is verified against the transcript, and your team makes the decision.

Is AI cheating detection legal in Europe?

It depends. The EU AI Act has generally banned AI emotion inference from biometric data in recruitment since February 2025, with exceptions for medical or safety uses. Recording, biometric processing and other monitoring need a case-specific legal assessment under GDPR. Ask your data protection lead before buying.

How do I stop candidates cheating with AI?

Publish a clear AI-use policy before the interview. Ask structured questions about the candidate’s own experience and follow up on specifics. Use job-specific scenarios. Verify at the next stage with a work sample or a hiring-manager conversation.

Is using AI to prepare for an interview cheating?

No. Rehearsing, researching and rewriting a CV with AI is preparation. Feeding answers live during the interview is the behaviour a policy should address. Say which one you mean, and say it before the interview.