AI Cheating in Job Interviews: What Happened, How Employers Responded, and Where the Line Is

In early 2025 a Columbia University student named Roy Lee said he had landed an Amazon internship with the help of a tool he built himself. Interview Coder sat on his screen during technical interviews, read the questions and suggested answers, and stayed invisible to the interviewer. Columbia took disciplinary action. Soon after, Lee and his co-founder had left the university and raised $5.3 million for a startup, Cluely, with the slogan “cheat on everything”. Two months later Andreessen Horowitz led a $15 million round.

After that story the problem was out in the open. Hidden AI in live interviews is now something recruiters say they see regularly, large employers have changed how they hire because of it, and a related problem, fake candidates using deepfakes, has reached FBI warnings. This guide collects the numbers and what employers did about them, and proposes a line between AI that helps and AI that deceives.

We make Realtimer, a real-time AI assistant for sales, property and recruiting calls, and our terms forbid using it to cheat in exams or job interviews. Figures below come from named surveys and news reports, linked at the end; where we rely on secondary coverage, we say so. Checked October 2026.

How common is AI cheating in interviews?

Nobody measures it directly, because the people doing it don’t report it. What exists are recruiter surveys, candidate surveys and vendor data, and they point the same way.

Figure	What it measures	Source
35%	of recruiters say they see AI used live during interviews	Greenhouse
91%	of recruiters say they have detected candidate deception	Greenhouse AI in Hiring report, via HR Tech Edge
39%	of candidates said they used AI during the application process	Gartner survey of 3,290 candidates
6%	of candidates admitted to interview fraud: posing as someone else, or having someone pose as them	Gartner survey of 3,000 candidates
1 in 4	candidate profiles worldwide could be fake by 2028, Gartner forecasts	Gartner, via HR Dive
63%	of job seekers have now been interviewed by an AI	Greenhouse 2026 Candidate AI Interview Report

The 35% from Greenhouse counts recruiters who say they see AI used live during interviews; it doesn’t say how many interviews involve it, or whether the use was hidden. The 91% is about deception of any kind, including exaggerated resumes. Coverage of the same Greenhouse report says 65% of hiring managers have caught applicants using AI dishonestly, including reading AI-generated scripts (32%) and appearing as deepfakes (18%). Those figures come from secondary reporting on the survey, not the report itself.

Vendors that sell detection publish bigger numbers. Fabric, an interview platform, says 38.5% of candidates in interviews it analysed between July 2025 and January 2026 were flagged for AI-assisted cheating behaviour. A company selling a fix has a reason to find a large problem, so read figures like that as a sign the problem is growing, and be careful with the exact size.

Candidates see it differently. In Greenhouse’s 2025 Workforce & Hiring Report, only 19% of US candidates said using AI in a live interview crosses the line.

Timeline: how we got here

When	What happened
2020 to 2022	Interviews move to video during the pandemic and mostly stay there. Hiring gets faster and cheaper, and the candidate’s screen becomes invisible to the interviewer.
September 2024	Deloitte UK says it will return to in-person final interviews for graduates and apprentices after the Financial Reporting Council flags cheating risks in fully online recruitment.
January 2025	The FBI updates its warning on North Korean IT workers and names AI and face-swapping in video interviews as one of their techniques.
Early 2025	Amazon’s internal guidelines on AI in interviews are reported by Business Insider. Anthropic’s application forms ask candidates not to use AI assistants.
April 2025	Roy Lee and Neel Shanmugam, the students behind Interview Coder, raise $5.3 million for Cluely.
June 2025	Andreessen Horowitz leads Cluely’s $15 million Series A. The FBI searches premises in 14 states linked to laptop farms used by North Korean workers.
June 2025 (Canva)	Canva tells engineering candidates it expects them to use AI coding tools in technical interviews. Meta’s test of AI-assisted coding interviews is reported weeks later.
July 2025	Gartner forecasts that one in four candidate profiles worldwide could be fake by 2028.
August 2025	Google says it will bring back at least one in-person interview round for some roles.
March 2026	Cluely’s CEO publicly retracts the $7 million revenue figure he gave TechCrunch the previous summer.
May 2026	Greenhouse reports that 63% of job seekers have been interviewed by an AI.

Why remote interviews made this possible

In a meeting room, the interviewer sees everything the candidate sees: the table, a glass of water, maybe a notebook. On a video call, the interviewer sees a face and whatever the candidate chooses to share. Everything else on the candidate’s desk, including a second screen, a phone and any window excluded from capture, is out of view.

For years that gap didn’t matter much, because there was nothing useful to put in it. Searching the web during a technical question was slow and obvious. What changed in 2023 and 2024 was speed: speech recognition that transcribes a question as it’s asked, and language models that can answer it in a few seconds. Put the two together, and anything an interviewer asks can be answered on a screen they can’t see before the candidate has finished saying “that’s a great question.”

Hiring processes also leaned harder on exactly the formats these tools are best at. Standard technical problems with known solutions, and behavioural questions with a recognisable shape (“tell me about a time you disagreed with your manager”), are the easiest things for a model to answer well. Interviews that depend on the candidate’s own history and judgement are harder to fake, and they’re also harder to run at scale, which is why many companies drifted away from them.

The Cluely story, from Interview Coder to “cheat on everything”

Interview Coder targeted one specific format: the remote coding interview, where a candidate solves an algorithm problem while an engineer watches over a video call. The tool read the problem and showed a solution in an overlay that screen sharing didn’t capture. Lee said he used it on interviews with large tech companies, including the Amazon internship. Amazon declined to comment on his case, telling TechCrunch that candidates must acknowledge they won’t use unauthorized tools.

The founders turned the controversy into marketing, and the April 2025 seed round of $5.3 million came from Abstract Ventures and Susa Ventures. In June 2025 Andreessen Horowitz led a $15 million Series A; investors not in the deal told TechCrunch the valuation was around $120 million.

Cluely’s manifesto, published on its website, opens the argument with one line: “We want to cheat on everything.” Its examples are sales calls, meetings and negotiations; it doesn’t mention interviews or exams. It says calculators, spellcheck and Google were all called cheating before they became normal, and that knowing the right question will matter more than memorising answers. It ends: “So, start cheating. Because when everyone does, no one is.” The site’s top menu also has a page called Undetectability.

The comparison works for everyday work, where everyone knows the tools exist and uses them. It breaks down in an assessment, which exists to measure what one person can do. Exams that allow calculators say so in the rules, and the ones that don’t still ban them. An interview works the same way: what matters is whether the person on the other side knows the AI is there.

The product then moved away from interviews, towards sales calls and meetings. In July 2025 Lee told TechCrunch that Cluely’s annual recurring revenue had doubled to about $7 million in a week. In March 2026 he retracted that on X, calling it the only blatantly dishonest thing he had said publicly.

Cluely wasn’t alone. Fabric’s data says the tools it caught most often were dedicated real-time prompters like Cluely and Interview Coder, followed by voice mode on a general chat assistant running on a second device.

How hidden interview assistants work, and why they’re hard to see

They all do three things. They listen to the call, either through the computer’s audio or a second device’s microphone. They turn the question into a prompt for a language model. And they show the answer where the candidate can read it while looking roughly at the camera: an overlay window that the operating system excludes from screen capture, or a phone propped below the screen.

Screen sharing doesn’t reveal them, because the window isn’t part of what’s shared. That’s the same technique that keeps a presenter’s notes private, and plenty of legitimate apps use it. Candidates give themselves away through behaviour: the pause before each answer, eyes that track text, answers that are fluent on the first question and vague on the follow-up.

What interviewers notice

No single sign proves anything; people pause, look away and read notes for innocent reasons. Interviewers who catch hidden assistance usually notice a pattern across the whole conversation. The signs below come from Amazon’s guidance to recruiters as reported, and from the FBI’s descriptions of fraudulent video interviews.

A steady delay before every answer, including simple ones, followed by fluent, complete sentences.
Eyes moving left to right as if reading, often slightly off the camera.
Typing while the question is being asked.
Polished first answers and vague follow-ups, especially on details only the candidate would know.
Answers that use the textbook vocabulary of a topic but none of the candidate’s own context: no names of systems, colleagues’ roles or numbers.
For identity fraud: audio that doesn’t match lip movements, and coughs or sneezes that don’t line up with the video, which the FBI has described.

Nervous candidates show some of these too. That’s why a suspicion calls for a better question first.

What employers did about it

Employers responded in three ways. Some wrote rules, some moved interviews back into the room, and some redesigned the interview itself.

Rules: Amazon and Anthropic

Amazon issued internal guidelines, reported by Business Insider in early 2025, telling recruiters to share the rules with candidates and warning that using generative AI in interviews without permission can get a candidate removed from the process. The guidance gives recruiters signs to watch for, including candidates who seem to read their answers or type while being asked questions. Amazon’s stated reason is that the tools give an unfair advantage and stop interviewers from seeing a candidate’s own skills.

Anthropic, which makes the Claude models, went further in its application forms: “please do not use AI assistants during the application process,” with a box to confirm. The same form encourages employees to use AI on the job. Anthropic’s stated reasons: it wants to see a candidate’s own interest in the company, and their communication skills without an AI in between.

Rooms: Google, Deloitte and others go back to in-person

Google brought back at least one in-person interview round for some roles in 2025. Sundar Pichai said that, with everyone working hybrid, it was worth having some fraction of interviews in person, and Google’s recruiting lead acknowledged that virtual interviews had sped up hiring but raised doubts about authenticity.

Deloitte UK moved earlier, and under pressure from a regulator. After the Financial Reporting Council flagged the risk of cheating in its fully online graduate recruitment, the firm returned to in-person final-stage interviews for graduate and apprenticeship programmes, while keeping early stages online. Reports in 2025 named Cisco and McKinsey among other employers reintroducing face-to-face meetings.

Redesign: interviews that assume AI exists

A third group stopped trying to keep AI out and changed what they test. Take-home tasks with AI explicitly allowed, followed by a live conversation about the choices the candidate made. Questions about the candidate’s own past work, with follow-ups that only someone who did it can answer. Pair programming where the interviewer watches how a person uses tools, not whether they do. These formats are harder to run, and they’re the ones hidden prompters help least with.

Canva made the switch publicly in June 2025. Its head of platforms, Simon Newton, published a post titled “Yes, you can use AI in our interviews. In fact, we insist.” Candidates for front-end, back-end and machine learning roles now solve open, practical problems with AI coding tools of their choice, and interviewers watch how they work with them. Newton said one reason was that candidates were increasingly using AI anyway and covert use was hard to detect; some Canva engineers worried at first that fundamentals would stop being tested.

Meta began testing the same idea in 2025. According to internal messages reported by 404 Media, it was developing a coding interview in which candidates have an AI assistant, arguing that it’s closer to how its engineers work and makes cheating with a hidden model harder. The reports described it as a limited trial.

The bigger threat: candidates who aren’t who they say they are

Reading answers off a hidden window is one problem. A candidate who isn’t the person on the CV is a worse one, and AI has made that easier too.

The FBI has warned repeatedly that North Korean IT workers apply for remote jobs at Western companies under false identities, and in January 2025 it named the use of AI and face-swapping during video interviews as one of their techniques. Its guidance tells employers to check identity documents carefully, verify employment and education directly with former employers and schools, and consider meeting candidates in person. In June 2025 the FBI searched premises in 14 US states linked to so-called laptop farms that let these workers appear to be based in the US.

The security company KnowBe4 described hiring a remote engineer who passed video interviews, background checks and reference calls, then turned out to be a North Korean worker using a stolen US identity and an AI-enhanced photo; it found out when he started loading malware onto his company laptop. Gartner analyst Jamie Kohn said candidate fraud creates cybersecurity risks that can be far more serious than a bad hire.

This is why in-person rounds and identity checks are coming back for remote roles, and it’s worth separating from the interview-prompter debate. A candidate who reads AI answers is misrepresenting their skills. A deepfake candidate is misrepresenting who they are.

The other side of the table: employers interview with AI too

Candidates point out, fairly, that employers are using AI on them. Greenhouse’s 2026 Candidate AI Interview Report found that 63% of job seekers had been interviewed by an AI, 13 points more than six months earlier. Resume screening, chatbot recruiters and one-way video interviews scored by software are common. Gartner found that only about a quarter of candidates trust AI to evaluate them fairly.

The difference, when it’s done properly, is disclosure. An employer that tells you an AI will ask the questions is using a tool openly. A candidate who hides a model behind the camera is presenting its answers as their own. Employers that run opaque AI screening and then ban candidates from using AI make that argument harder to win, and the regulation below is pushing them towards transparency.

Where the line is between help and deception

A useful test: is someone being misled about what they’re evaluating? If the interviewer believes they’re hearing your knowledge and they’re hearing a model, that’s deception. If everyone knows what’s in the room, or nobody is being evaluated on the thing the AI provides, it isn’t.

Situation	Verdict	Why
Practising with AI before the interview: mock questions, feedback on your answers	Fine	Nobody is misled. It’s rehearsal, like practising with a friend.
Using AI to tighten the wording of your resume	Fine, if every claim is true	The facts are yours; you can defend each line in the interview.
AI that adds experience or skills you don’t have	Deception	The resume describes someone else.
AI on a take-home task where the employer allows it	Fine	The rules were set and you followed them. Say which tools you used.
Hidden AI feeding you answers during a live interview	Deception	The interviewer is assessing you, and is hearing a model.
A second phone running voice mode during a video interview	Deception	The same thing with a different device.
Someone else, or a deepfake, attending the interview for you	Fraud	Identity fraud, and in some cases a security threat.
An employer interviewing you with an AI, and saying so	Fine, with disclosure	Both sides know who, or what, is asking the questions.
A recruiter using an assistant for company facts and notes, disclosed where required	Fine	It’s the employer’s own information, and nobody is being evaluated on it.

A recruiter who uses an assistant to answer a candidate’s question about the salary band or the on-call rota, from the company’s own handbook, isn’t misleading anyone about their own skills; the candidate wants the company’s facts and gets them faster. The same software used by the candidate to answer the recruiter’s questions would be on the wrong side of the line.

For candidates: why it backfires even when it works

Suppose the hidden assistant gets you through. You now start a job you were hired for on the strength of someone else’s answers, with a manager who expects the person from the interview. Probation periods exist for exactly this. In technical roles the gap shows within weeks; in client-facing roles, within the first difficult call.

It’s also riskier than it looks during the interview. Interviewers trained to look for it notice the pauses and the reading, and the follow-up question is hard to fake: “You said you migrated the billing service. What broke first?” A model can write a convincing first answer; it doesn’t know what happened at your last job.

The better use of AI is before the call. Practise with mock questions, ask for blunt feedback on your answers, research the company, and make sure your resume says what you actually did. Our guide to free ATS resume checkers covers the resume side. Then walk into the interview with your own answers, and if you don’t know something, say so and say how you’d find out; interviewers often rate that higher than a polished guess.

For recruiters: how to run interviews when AI is everywhere

These steps are what employers in this guide have done, plus what follows from them.

Say your AI policy before the interview, in writing: what’s allowed in preparation, in take-home work and in the live conversation. Amazon’s rules work partly because candidates confirm them.
Ask about specifics from the candidate’s own experience, then follow up twice. Depth on their own story is the hardest thing to prompt.
Use work samples with stated rules, then discuss the choices live.
For remote roles, verify identity: documents, direct checks with former employers, and a conversation with the camera on, where you can ask unexpected questions.
Consider one live or in-person round for senior or high-risk roles, as Google and Deloitte UK did.
Don’t rely on AI detectors alone. They produce false positives, and a wrong flag can end a real candidate’s application; non-native speakers are the group most often mentioned.
If you use AI in your own process, disclose it. It’s increasingly a legal requirement, and it’s hard to ask candidates for honesty you don’t offer.

Interview questions that are hard to fake

The best defence is a question whose answer only exists in the candidate’s head. These work in most professional interviews, technical or not.

“Walk me through the last project on your CV. What was the hardest decision, and who disagreed with you?” Then ask what happened next, twice.
“What number did you track to know it was working? What was it before and after?” People who did the work remember the number, or remember why there wasn’t one.
“What would you do differently if you started it again?” A real answer is specific and slightly uncomfortable.
“Explain it to me as if I were new on your team.” Changing the audience in the middle of an answer is hard to prompt fast.
For technical roles: change the constraints of a problem halfway through and ask what breaks.
“Tell me about something on your CV you’d remove now.” There’s no standard answer for a model to reach for.

These questions also make better interviews in general. They tell you how someone thinks, which is what you wanted to know before anyone had heard of Cluely.

If you suspect a candidate is using AI

Handle it the way you’d handle any doubt in an interview: with more information, consistently, and in writing. Move to the questions above. If your policy allows it, ask the candidate to share their entire screen or to show their desk briefly, and explain that you ask everyone. Don’t accuse anyone on the call; a wrong accusation is unfair to the candidate and a legal risk for you.

After the call, write down what you observed, not what you concluded: “paused about five seconds before each answer; could not describe the database migration listed on the CV.” If the role matters, add a live or in-person round. And apply the same steps to every candidate, so the decision rests on the process rather than on one interviewer’s instinct.

What the law says

There’s no general law against a candidate using AI in an interview; it’s an employer’s rule, enforced by withdrawing an offer. The law is mostly about employers’ use of AI, and about recording.

In New York City, Local Law 144 requires employers using automated employment decision tools to run bias audits and notify candidates. The EU AI Act classifies AI systems used for recruitment and evaluating candidates as high-risk, with obligations on transparency, human oversight and data quality. Separately, many countries and US states require consent before a call is recorded or transcribed, which applies to note-taking tools on either side of an interview. Check the rules where you and the candidate are; this is a summary, not legal advice.

If you use AI interviewers, how to keep it fair

Employers that ask candidates to keep AI out of interviews are on firmer ground when their own use of AI is open. Tell candidates in the invitation when an AI will ask questions or score answers, and what a human reviews. Give them a way to ask for a conversation with a person. Check the system for bias, which New York City already requires for automated decision tools. And keep the final decision with a person who can explain it, which the EU AI Act expects for high-risk systems such as recruitment.

Candidates notice the double standard quickly. Only about a quarter trust AI to evaluate them fairly, according to Gartner, and every unexplained rejection by a system they couldn’t see makes the next hidden assistant easier for them to justify.

For recruiters: help on your side of the call

If you run interviews and want quick, accurate answers to candidates’ questions about the role, salary band and next steps, from your own documents, with notes after the call, see Realtimer for recruiters.

AI in job interviews: FAQ

Is it cheating to use AI in a job interview?

Using hidden AI to answer questions in a live interview is deception: the employer is assessing your knowledge and communication, and is hearing a model instead. Preparing with AI beforehand, or using it where the employer explicitly allows it, is not. If you’re unsure, ask the recruiter what’s allowed before the interview.

Can interviewers tell if you’re using AI?

Often, yes. Amazon’s guidance to recruiters, as reported by Business Insider, lists signs such as reading answers instead of speaking naturally, typing while being asked questions and eyes moving across the screen. Follow-up questions about specifics are the most reliable test: a model can produce a polished first answer, but it can’t remember what you did at your last job.

What happens if you’re caught using AI in an interview?

At companies with explicit rules, you can be removed from the process; Amazon’s guidelines say so. If it’s discovered after hiring, it can lead to dismissal, and in regulated professions it can lead to disciplinary action.

Which companies have returned to in-person interviews?

Google has said it will bring back at least one in-person round for some roles. Deloitte UK returned to in-person final interviews for graduate and apprenticeship programmes after a regulator raised cheating risks in fully online recruitment. Reports also name Cisco and McKinsey. Most companies keep early rounds online.

Is it legal for companies to interview candidates with AI?

Generally yes, with rules. New York City requires bias audits and notices for automated employment decision tools. The EU AI Act classes AI used in recruitment and candidate evaluation as high-risk, with obligations for transparency and human oversight. Many jurisdictions also require consent to record a call.

What is Cluely?

Cluely is a startup that grew out of Interview Coder, a tool its founders built at Columbia University to help candidates in technical interviews. It raised a $5.3 million seed round in April 2025 with a “cheat on everything” message, then a $15 million Series A led by Andreessen Horowitz in June 2025, and has since moved towards sales and meetings. In March 2026 its CEO publicly retracted a revenue figure he had given to TechCrunch.

What are deepfake job candidates?

Applicants who use face-swapping or AI-generated video to appear as someone else in a video interview. The FBI has warned that North Korean IT workers use these techniques to get remote jobs at Western companies, and security firm KnowBe4 has described hiring one who passed video interviews and background checks.

How should recruiters adapt their interviews?

Say clearly what AI use is allowed. Ask follow-up questions about specifics from the candidate’s own experience. Use work samples with stated rules. Verify identity for remote roles. Consider one live or in-person round for senior roles. Don’t rely on AI detectors alone; they produce false positives.

Should candidates disclose that they used AI to write their resume?

Not usually needed, as long as everything in it is true and you can talk about every line. If an application form asks directly, answer honestly. Anthropic, for example, has asked applicants not to use AI assistants in its application process and to confirm they haven’t.