AI cheating in interviews: the 2026 state of play (and what to actually do about it)

AI cheating in interviews is mainstream in 2026. Interview Coder advertises over 150,000 users at $799 lifetime access and a claimed 65% success rate. Final Round AI markets itself with more than 10 million users. Cluely, LockedIn AI, Linkjob, and Parakeet AI all operate openly as real-time, candidate-side, "invisible to screen sharing" overlays.

The question is no longer whether candidates are using these tools. It is what your hiring system does with that fact. The serious answer is one of three responses, and the wrong one is "more detection."

The calm version of this article is that the equilibrium has shifted, the headlines are catching up, and there are three coherent ways hiring leaders are responding. Detection-only is the most common and the worst. Redesign-and-allow (the Canva model) and redesign-and-replace (the conversational AI interview model) both work for different contexts. The point of this piece is to help you pick which one fits your situation, not to sell you the third.

Key Takeaways

• AI cheating in interviews is no longer a fringe behaviour. Real-time AI overlays have millions of users between them and are marketed openly as invisible to screen sharing.

• Detection-only is a losing strategy: the escalation dynamic favours the cheat tools, false positives cost you good candidates, and detection-as-AEDT triggers regulatory exposure under NYC LL144 and the EU AI Act.

• The two responses that actually scale are: redesign the interview to allow AI use openly with rules (the Canva model), or replace the cheatable round with a conversational AI interview whose follow-up structure makes cheating irrelevant.

• The "cheaters lack competence" framing is wrong and damaging. Many strong engineers use AI well. Conflating legitimate AI use with cheating is a category error that loses you good hires.

• The right question for hiring leaders this quarter is not "how do we catch them" but "what hiring system gives us a real signal in a world where AI assistance is mainstream."

Where the market actually is in mid-2026

Three data points anchor the state of play.

The candidate-side tools are not hiding. Interview Coder publishes its user count (over 150,000 in 2026) and its success-rate claim (65% on coding rounds) on its own homepage. Final Round AI claims more than 10 million users across coding, behavioural, and system design rounds. Cluely and LockedIn AI market themselves as invisible to screen sharing on Zoom, Google Meet, and Microsoft Teams.

ChatGPT in job interviews is the manual baseline version of the same behaviour: a candidate typing the question into a chat tab during a remote interview. It has been around longer than the overlays and is in many ways the entry point to the category.

AI cheating in coding interviews specifically is the highest-prevalence sub-category because the cheat-tool quality on standard coding questions is highest. The whole category is at the "publicly marketed SaaS product" stage, not the "underground forum" stage.

Hiring teams are starting to react, unevenly. Google's Sundar Pichai reportedly told an internal town hall in February 2025 that the company would return to in-person interviews for some roles to address the rise in candidates using AI in interviews, and Deloitte reinstated in-person interviews for parts of its UK graduate program. Specific founder anecdotes circulating in 2026 include one technology startup reporting that "more than 50%" of candidates cheated on a recorded coding challenge. There is no peer-reviewed industry-wide prevalence number on AI interview cheating yet; treat the anecdotes as directional and the trend as real.

The asymmetry is real. Candidate-side tools improve every quarter. Detection tooling improves more slowly because the relevant signals (eye-flicker, latency, voice mismatch) are hard to extract reliably at scale and are easy for the cheat tools to design around. Pure detection arms races favour the offence. That is not a moral claim; it is the structural shape of the problem.

The honest summary is that the equilibrium has shifted. A 2023 HireVue survey found more than 85% of HR leaders worried about candidates cheating during virtual interviews. The 2026 reality is that the worry was justified, the tooling has caught up to it, and the hiring systems most companies still run were not built for this environment.

The three responses hiring teams are choosing between

There are exactly three coherent responses to this dynamic. Everything else is some confused mix that performs worse than any of the three.

Response 1: detection-only. Add a proctoring layer, an AI cheating detector, or a manual review process to your existing interview rounds. Keep the rounds themselves the same. Try to catch the cheaters.

Response 2: redesign and allow. Change the rounds so AI use is expected and the questions are designed for someone using AI. This is the Canva engineering team's published policy: backend, frontend, and ML candidates are expected to use Copilot, Cursor, and Claude during coding interviews. The questions changed from algorithmic puzzles to complex, ambiguous, real-world problems where AI is part of the work.

Response 3: redesign and replace. Replace the cheatable rounds with a structured, conversational interview that prices the cheating dynamic in. Live conversation, hard follow-ups, scoring against a rubric the candidate cannot prepare answers for. This is what AI interview platforms (including Expert Hire's) are designed to do; the same shape works with a human interviewer following a structured rubric.

Each response fits a different context. Picking the wrong one is most companies' actual problem in 2026.

Why detection-only is a losing strategy

Detection-only is the most-attempted and least-effective response, for three reasons.

The escalation dynamic favours the cheat tools. Detection-only is an arms race. The detector vendor updates its signals; the cheat-tool vendor reads the update, ships a patch, and stays ahead. Final Round AI and Cluely have shipped quarterly updates designed specifically to evade common detection signals. You are not paying for a permanent solution; you are paying for a subscription to a treadmill.

False positives cost you the candidates you most want to hire. Detection signals (response latency, eye flicker, voice mismatch) correlate weakly with cheating and strongly with normal interview behaviour: a nervous senior candidate, a non-native speaker thinking carefully before answering, a candidate using assistive technology, or a candidate who happens to have a notebook off-camera. The 5% false-positive rate that sounds tolerable on a slide deck means dozens of strong candidates per quarter incorrectly flagged. Some never come back.

Detection systems are AEDTs and trigger regulatory exposure. Under NYC Local Law 144, a tool used to substantially assist a hiring decision is an Automated Employment Decision Tool and requires a published bias audit and candidate notice. The same logic applies under the EU AI Act for hiring AI in EU markets, and under the Illinois AI Video Interview Act for video interviews of Illinois residents. A cheating detector that materially shapes whether a candidate advances is exactly the tool these laws were written to cover. The procurement and legal cost of running it correctly is not trivial.

Detection is useful as a layer inside a redesigned interview. It is not useful as the whole strategy. Treating it as the strategy is what most companies are getting wrong this year.

What the two redesign responses look like in practice

The two real options each have a published, public example. The Canva engineering team has written about the allow-and-redesign model. The conversational AI interview model is what platforms like Expert Hire are built around.

The Canva model: allow it openly, redesign around it

The Canva engineering team published a policy in 2025 that other engineering orgs are now copying. The summary: AI use during the coding interview is not just allowed, it is expected. The interview questions changed accordingly.

Specifically: Canva replaced the computer-science-fundamentals screening (the type of question a candidate-side AI tool trivially solves) with an "AI-Assisted Coding" interview built around complex, ambiguous, real-world problems. The example they cite is "build a control system for managing aircraft takeoffs and landings at a busy airport." The candidate uses Copilot, Cursor, or Claude during the round. The interviewer scores judgment, not raw coding speed.

The bar Canva sets for candidates: ask thoughtful clarifying questions about product requirements; use AI strategically for subtasks while maintaining solution control; critically review the generated code; debug effectively when issues arise.

The reasoning is structural. As the Canva team notes, traditional interview formats "didn't reflect actual work conditions." Nearly half their engineers use AI coding assistants daily. Assessing candidates without those tools produces no meaningful signal about real job performance. So the cheating problem is solved at its root: there is no advantage to hiding AI use because AI use is the point.

The model fits engineering orgs whose real day-to-day involves heavy AI assistance and whose interview process can be redesigned without months of compliance review. It is harder to apply for regulated industries (finance, healthcare, defence) or for jobs where AI use in the real role is restricted or audited.

The conversational AI interview model

The third response is to replace the cheatable round with a structured conversation. Not a recorded one-way video; a live, voice-to-voice conversation against a published rubric, with follow-ups generated based on what the candidate actually said.

Why this neutralises cheat tools: candidate-side overlays handle the prepared question reasonably well and break on the follow-up. The overlay can transcribe the question, hand it to a model, and surface a suggested answer in three to four seconds.

What it cannot do is help the candidate answer "why did you choose that approach over X" when the candidate hasn't actually thought through X. The model only surfaced one answer, and the latency to surface a second collapses the conversation. Interviews structured around hard follow-ups produce signal in this environment. Interviews structured around a single set-piece question do not.

We built Expert Hire's AI interview platform for exactly this dynamic. The scoring approach is published openly on the methodology page; the artefact each interview produces (rubric, transcript, reasoning per criterion) is exactly the artefact a hiring manager needs to make a decision that survives an audit. The same conversational, follow-up-heavy structure works with a human interviewer trained to push past the first answer.

This is the bias we have, and we are being explicit about it. We built the product the way we did because, of the three responses, this one scales the best across engineering, business, and operations hiring.

The regulatory exposure is also the smallest: a structured-interview tool with a published bias audit is a much simpler compliance case than a covert detection layer. The Canva model is fine; the conversational AI interview model is fine; detection-only is not.

The "cheaters lack competence" position is wrong

A popular take in the practitioner blogs of 2025 and 2026 is that engineers who use AI assistance during interviews lack the underlying skill, and that hiring leaders should treat the use itself as evidence of incompetence. The take is wrong and it is damaging.

It is wrong because plenty of strong engineers use AI assistance. The Canva data point (about half of Canva engineers use AI coding assistants daily) is broadly representative of the industry. Treating AI use as evidence of weakness conflates a normal work tool with deception. That conflation costs you the engineers whose actual problem is that the interview banned a tool their day job runs on.

It is damaging because it pushes hiring leaders toward detection-only, which (as covered above) is the worst response. The frame "catch the cheaters because they can't do the work" leads to dashboards full of false positives and a candidate experience that loses you good hires.

The accurate distinction is between AI use (transparent, expected, scored on judgment) and AI deception (covert, designed to fake competence the candidate doesn't have). The first is fine and increasingly normal. The second is the actual problem. Conflating them buys you the worst of both worlds: an interview process that punishes legitimate workflow and still doesn't filter out the candidates who can't reason past a model's first answer.

The legal layer (briefly)

A short note on the law, because it shapes what responses are viable.

NYC Local Law 144 treats any tool that substantially assists a hiring decision as an Automated Employment Decision Tool. AI cheating detectors, AI interview scoring systems, and AI screening tools all qualify. Each requires a published annual bias audit and a candidate notice. If you are deploying any of these in a New York City hiring process, the compliance cost is real.

The EU AI Act classifies hiring AI as high-risk. The compliance requirements (conformity assessments, technical documentation, risk management, human oversight) apply whether the AI is a scoring tool or a detection tool. The phase-in dates run through 2026 and 2027; the prep is real.

The Illinois AI Video Interview Act requires candidate consent and disclosure when AI analyses video interviews of Illinois residents. The act is older than the cheat-tool category but applies cleanly to it.

For a fuller picture of the rules across US states, the state-by-state map is the right starting point.

The legal picture matters here because it changes the cost of the three responses. Detection-only has the highest regulatory exposure (the detector is squarely an AEDT). Redesign-and-allow has the least (no AI is making the hiring decision). Redesign-and-replace sits in the middle (a structured AI interview is an AEDT, but a well-documented one is the simplest case to comply with).

A three-question decision framework for hiring leaders this quarter

Three questions to bring to your next interview-process review.

One: what does AI use actually look like in the role we are hiring for? If the answer is "daily, deeply, all over the workflow," the Canva model fits. If the answer is "limited or audited," the conversational AI interview model is a better fit. If the answer is "you genuinely don't know yet," go find out before designing the interview.

Two: what is the signal each round is supposed to surface? A screening round whose only signal is "can this candidate write a working function under time pressure" is the easiest round in the world for a cheat tool to defeat. Replace it with a round that surfaces judgment, problem decomposition, or trade-off reasoning. If the round can be done well by an overlay, it is the wrong round.

Three: where does your regulatory exposure live? If you hire in New York City, Illinois, the EU, or any combination, detection-only is the highest-cost option you have. The two redesign responses scale much better as you add jurisdictions.

These questions are deliberately not "which vendor should I buy." The point of the framework is to make the system-level decision first. The vendor question is downstream.

Frequently asked questions

How widespread is AI cheating in interviews in 2026? There is no peer-reviewed industry-wide prevalence number. Specific founder reports cite more than 50% of candidates cheating on recorded coding challenges in some funnels; published user-count claims from cheat-tool vendors (Interview Coder over 150,000, Final Round AI more than 10 million) suggest the category is mainstream rather than fringe. Treat the situation as "common, growing, and you should plan for it" rather than as "exactly X% of candidates."

What tools are candidates actually using? Real-time overlays that listen to the interviewer through system audio and display suggested answers on a transparent layer the candidate reads while facing the camera. Specifically: Interview Coder for coding rounds, Final Round AI across coding and behavioural, Cluely and LockedIn AI for general-purpose live-interview assistance, Parakeet AI and Linkjob AI as cheaper alternatives. Several explicitly market themselves as "invisible to screen sharing."

Should you ban AI in interviews? Sometimes. The right answer depends on whether AI use is normal in the actual role and on what signal the round is meant to surface. Banning AI in a coding round for a role whose day-to-day involves daily AI coding assistance gives you no useful signal. Allowing AI openly with a redesigned question gives you a much better signal. The right policy is per-round and grounded in role, stage, regulatory context, and the signal each round is meant to produce. There is rarely a single answer that fits an entire hiring loop.

Can you actually detect AI cheating? You can detect some tells some of the time. Response latency that doesn't match question difficulty, eye-movement patterns locked to a single off-camera region, voice and confidence step-changes between the prepared answer and the unprepared follow-up, and the "burst-then-defend" coding pattern (a clean function appears in 30 seconds, the candidate then cannot explain why) are the tells our team and others have flagged consistently. The detection works inside a well-designed interview; it does not work as a standalone strategy.

Will AI cheating in interviews get worse? The candidate-side category will keep improving and keep being marketed openly. The right framing is not "will this get worse" but "is the hiring system you're running going to produce a real signal in a world where AI assistance is mainstream." Companies redesigning their rounds (response 2 or 3) are pulling ahead. Companies layering more detection on the same rounds (response 1) are falling behind.

What's the legal exposure if we use AI cheating detection? Depends on jurisdiction. In New York City, an AI cheating detector that materially shapes hiring decisions is an AEDT under Local Law 144 and requires a published bias audit and a candidate notice. In Illinois, AI analysis of video interviews requires candidate consent under the AI Video Interview Act. In the EU, hiring AI is high-risk under the EU AI Act and triggers conformity-assessment obligations. None of this is fatal; all of it is real cost.

The point isn't to catch every cheater

The point of any response to AI cheating in interviews is to make sure the signal your hiring system produces is real. Detection-only doesn't do that. It produces a list of suspects, some real and some not, and leaves the underlying interview design intact.

The two responses that work are structural. Redesign the round so AI use is allowed openly and the question rewards judgment, or replace the round with a structured conversation whose follow-ups make cheating irrelevant.

Both are real options. Both scale. The hiring leaders who will look smart in two years are the ones picking between them now.

If you want to see how the conversational AI interview model handles this in practice, open Expert Hire's AI interview platform, look at a sample scorecard with the rubric and the reasoning per criterion, and decide for yourself whether the follow-up structure addresses the dynamic this article is about. That is the artefact that does the work, on the hiring side and on the audit side.

• By TK, Growth at Expert Hire. Last updated June 11, 2026.