Is it cheating to use AI in a technical interview?

It depends on disclosure, not the tool itself. A candidate openly using an AI agent in a workspace built for it is doing the job the way it's actually done. A candidate running a hidden overlay against instructions, in an interview that never told them AI was fair game, is misrepresenting their own ability — that's the cheating.

How common is AI cheating in interviews?

Fabric's analysis of 19,368 interviews found AI-assistance signals in 38.5% of them overall, rising to roughly 48% in technical roles specifically. Separately, Blind's survey of 3,617 verified professionals found 20% had secretly used AI in an interview and 55% called it the new norm.

Can interviewers actually detect tools like Cluely or Interview Coder?

Not reliably by watching the shared screen — these tools render an overlay beneath what screen-sharing software captures, by design. What's harder to fake is the behavior around it: answers with no visible reasoning trail, timing that doesn't match difficulty, or code that appears fully formed with no intermediate attempts.

What's the alternative to banning AI and hoping nobody uses it?

Give candidates a real, sanctioned AI workspace instead of an unmonitored one, and grade how they use it — the prompts, the verification, the moments they override a wrong suggestion. That turns the exact behavior you couldn't see in a ban-and-hope interview into the thing you're scoring.

Hiring

AI cheating in technical interviews, explained

Tarunpreet Singh — Co-founder, DeftBench

7 September 2026

AI cheating in technical interviews means a candidate using an AI tool to produce answers they present as their own, without the interviewer's knowledge or consent — not simply using AI at all. The distinction matters because the tools involved (silent overlays like Cluely and Interview Coder, voice-mode LLMs listening in the background) are now common enough that treating every AI-assisted interview as a data point in an arms race, rather than confronting why candidates feel they need to hide it, misses what's actually happening.

How common is AI cheating in technical interviews?

More common than most hiring teams assume, and rising. Fabric, which builds interview-integrity detection tooling, analyzed 19,368 interviews and found AI-assistance signals in 38.5% of them overall — and in roughly 48% of purely technical interviews specifically, more than double the rate in sales roles. Separately, Blind's survey of 3,617 verified professionals (conducted April 2025) found 20% admitted to secretly using AI during an interview, and 55% agreed it had become the new norm. Both are self-reported or vendor-measured numbers, not a controlled study — but they agree on direction: this isn't a fringe behavior, and it isn't going away.

What tools do candidates actually use to cheat with AI?

Per Fabric's breakdown of the same dataset, dedicated cheating assistants — Cluely and Interview Coder are the named examples — account for about 45% of detected cases. Voice-mode general-purpose LLMs (a candidate quietly talking to ChatGPT or similar) make up another 34%. The rest splits between low-tech methods like tab-switching or a second screen (18%) and live help from another person off-camera (3%).

The dedicated tools are the harder problem: they're built to render an answer as a transparent overlay that sits beneath the layer a screen-share captures, so the interviewer sees a clean shared screen while the candidate reads a generated answer off it in real time. That's a detection problem no amount of "please don't use AI" in the interview invite solves.

Is it actually cheating, or the new normal?

The honest line isn't AI use itself — it's whether the AI is enhancing a real capability or fabricating one that isn't there. A candidate who uses AI to organize their own genuine experience into a clearer answer is doing something closer to using notes in an open-book exam. A candidate who has an agent solve the coding problem while they read the solution off a hidden screen has produced no evidence of their own ability at all — the interview measured the model, not the candidate.

Banning AI outright collapses that distinction and creates the exact blind spot the numbers above describe: candidates are told not to use it, use it anyway, and the interviewer has no way to tell which candidates are doing which version of that.

Why detection and proctoring don't solve this on their own

Overlay tools are built to beat screen-share.If the rendering happens below what the conferencing software captures, watching the shared screen harder doesn't help.
Behavioral signals are noisy. Long pauses, halting speech, or re-reading a question can look like consulting a hidden tool — or like anxiety, a language barrier, or a slow connection. Calibrated wrong, detection tooling punishes honest candidates as often as it catches dishonest ones.
It's a permanent arms race. Every detection method (keystroke timing, eye-tracking, similarity scoring) gets a countermeasure within months. Chasing the tool is a treadmill; nothing about the interview design changes.

What actually works: make AI use part of the assessment, not a violation of it

If the job the candidate is interviewing for involves using AI tools daily, the fix isn't a better lie detector — it's removing the reason to lie. Give candidates a sanctioned AI workspace as the default, tell them plainly that using it is expected, and evaluate the parts that a hidden overlay can't fake: how they framed the problem, the specificity of what they asked the agent for, whether they verified what it produced, and whether they caught it when it was confidently wrong. A candidate can copy an answer from a hidden tool. They can't fake a session-long track record of good judgment about when to trust an agent and when to override it — not without that judgment actually existing.

We cover the mechanics of running that kind of interview — what to measure and how to structure the session — in how to interview engineers who use AI coding agents.

Where DeftBench fits

DeftBenchis built around this shift: candidates get a real agent in a live browser IDE instead of a banned tool they have to hide, and the session is captured as telemetry with human and agent actions attributed separately — so the evaluation is evidence-linked instead of a guess about what happened off-screen. If you're comparing this approach to an AI-assisted coding test with a bolt-on proctoring layer, see the DeftBench vs CoderPad comparison, or talk to us about a pilot.