AI Interview Copilots: Where Help Ends and Cheating Starts

In February 2025, Columbia sophomore Chungin “Roy” Lee posted a YouTube video of himself passing an Amazon interview with an app called Interview Coder open on a second screen. The app screenshotted the coding question the interviewer had just given him, ran it through a model, and handed him a working solution he read back as his own. Amazon had the video pulled for copyright infringement. Columbia opened a disciplinary case. Lee, 21, turned the episode into a startup that had raised $5.3 million in seed funding by that April.

So: is using an AI interview copilot cheating? Most people, including plenty of engineers who use Copilot at their actual jobs every day, would answer yes for the Lee case without much hesitation. LastRoundAI sells a version of the same category of tool: software that listens to a live interview and surfaces answers in real time. Which means this isn’t a neutral question for us to answer, and pretending otherwise would be its own kind of dishonesty. Here’s where we actually land, and why the AI interview assistant ethics question is narrower than it sounds once you separate two things that keep getting treated as one.

What Roy Lee Actually Did, and Why It Wasn’t a Gray Area

Lee’s app worked by screenshotting interview questions without triggering screen-share detection, then generating a solution live. By his own account to NBC News, he used it to land internship offers from Amazon, Capital One, Meta, and TikTok. He turned down Amazon’s offer himself. Several of the other companies rescinded theirs once he went public about how he’d gotten them. Columbia suspended him for a year. Interview Coder, meanwhile, grew 50% week over week and was pulling in $170,000 a month in subscriptions within weeks of the video going viral, figures Lee gave NBC directly.

Amazon’s on-record response is about as close to a bright line as a company gets. Spokesperson Margaret Callahan told NBC that candidates “must also acknowledge that they won’t use unauthorized tools, such as generative AI technology, to help them during the interview or assessment process.” Worth sitting with why that line exists: the round Lee gamed is designed to check whether a candidate can write working code under time pressure, unaided. That’s the entire point of it. Routing the answer through a hidden model doesn’t clear that bar. It replaces the bar with a different question nobody asked (can a language model solve this) and reports the result as if it answered the first one.

CNBC reported that Google CEO Sundar Pichai told hiring managers at a February 2025 town hall to consider going back to in-person interviews, specifically because of tools like Lee’s, just months after telling investors that AI was already writing more than a quarter of Google’s new code. That’s a real tension, not necessarily hypocrisy: reviewed, tested production code is a different animal from a timed, unaided live check. But it’s worth naming plainly rather than skating past it.

The Distinction That Actually Matters

Every interview round is a proxy for something the employer can’t observe directly: whether you can actually do the job. Some rounds try to measure that head-on. Write this function, right now, with nothing but your own head. Other rounds measure something adjacent. Can you talk clearly about work you’ve already done. Can you structure an answer instead of rambling for four minutes. Can you stay coherent when you’re nervous. Those aren’t the same test, and a copilot does something completely different to each one.

If the round is checking a capability you don’t have, and the tool supplies that capability in the moment, you’re lying about what you can do. That’s the Lee case exactly. It’s also true for using a copilot to generate original code in a live pairing round meant to observe your actual process, or to fabricate a systems-design answer for an architecture you’ve genuinely never touched. The tool isn’t helping you demonstrate a skill. It’s standing in for a skill you’ll need on day one and won’t have.

If the round is checking access, communication, or composure, that’s a different animal. “Tell me about a time you disagreed with your manager” isn’t asking whether you can invent a plausible story on the spot. It’s asking you to retrieve something true from memory and say it clearly under mild panic. If your mind goes blank two minutes into a panel round, which happens to most candidates at some point, a nudge that says “you mentioned a scheduling conflict on the Meridian project, use that one” isn’t inventing the conflict. It already happened. The tool there is closer to a teleprompter than a ghostwriter.

I don’t think this line is clean everywhere. A live system-design round that’s really testing tradeoffs and communication rather than raw novelty sits somewhere in between, and I don’t have a confident rule for exactly where it tips over. My honest guess is that it depends on whether the interviewer would still want to hire you if they found out mid-interview, which is a strange standard to apply, but it might be the real one underneath all the others.

Same Tool, Four Rounds, Four Different Answers

Where This Gets Uncomfortable for a Company Selling the Tool

LastRoundAI’s own marketing for the AI interview copilot leans on words like invisible and undetectable, phrasing that sits closer to how Lee described Interview Coder to CNBC than we’d probably like. That’s a fair criticism, and I’m not going to pretend otherwise. The distinction we’d argue for is use case, not mechanism: the same hidden overlay that’s indefensible on a live coding round built to test raw ability is, in our view, fine on a behavioral round built to test composure. But “the mechanism looks identical from the interviewer’s side” is a real problem with that argument, not a footnote to wave off. We’re not neutral on AI interview assistant ethics. We sell one of the tools in question. Weigh that before you take our word for anything in this post.

The honest answer is that the tool itself doesn’t know which kind of round you’re in. You do. That puts more responsibility on the person using it than most AI-tool marketing, ours included, wants to admit up front.

Is It Actually Against the Rules, or Just an Opinion?

It depends entirely on the employer, and there’s no single industry-wide rule yet. Anthropic’s own published guidance to candidates says, in its exact wording, “This is all you, no AI assistance” for live interviews, while explicitly welcoming AI help drafting a resume or preparing beforehand: “use it for preparation and refinement,” is how the company puts it, adding that it wants to “see your thinking in action” once the interview starts. Amazon’s stance, per its own spokesperson, bans unauthorized tools during interviews and assessments outright, with no carve-out mentioned for prep work. Neither company treats this as a gray area internally, even where the underlying ethics genuinely are gray to an outside observer.

Lee’s own argument, given to CNBC, is that the format itself is the real problem: “It doesn’t make sense to have an interview format that assumes you don’t have the use of AI.” That’s not a crazy position. LeetCode-style rounds are a famously poor proxy for day-to-day engineering work, and reasonable people have argued this for years, long before Lee showed up. But being right about the format’s flaws doesn’t retroactively make gaming it something other than what it is. Three companies still rescinded offers over exactly this. Being correct about a rule’s fairness has never been a defense against having broken it.

Where We Actually Land

Use an AI interview copilot to survive a bad night’s sleep, a blanked-out memory, or a panel moving faster than your nerves can keep up with. Don’t use it to manufacture a skill you don’t have and will be expected to produce alone starting your first day. And if a specific employer has told you, in writing, not to use one at all, a growing number now do, that instruction settles the question regardless of which side of our own argument you land on.

It’s also why we ship Mock Interviews and Interview Copilot as two separate products rather than folding them into one. Mock Interviews is built for generating your own answers out loud and getting scored on structure and delivery, no assistance, because that’s the tool actually meant to build the skill. Interview Copilot exists for the live moment itself. Treating the two as interchangeable would undercut everything argued above.

What people actually use it for

We can put a number on the assumption underneath this whole argument. Across 1,393 interview sessions configured on LastRound between January 2025 and July 2026, 721 were set up as live sessions and 672 as mock practice. Close to an even split.

We expected live to dominate by a wide margin. It doesn’t. Roughly half of all configured sessions are somebody rehearsing against nobody, which is the use case no one argues about and no one writes think-pieces on either.

One more from the same set, and this one is genuinely odd: 1,374 of those 1,393 sessions were set to intermediate difficulty. Three people picked beginner. Sixteen picked expert. We shipped a difficulty selector that essentially nobody moves, which says something about the setting and possibly more about the selector.

FAQ

Is it cheating to use AI in interviews for every type of question?

No. It depends on what the specific round is testing. Using AI to generate an answer to a live technical question meant to prove you can code unaided is faking a capability. Using it to help structure a true behavioral answer under nerves is closer to assistance than fabrication. The two get lumped together constantly, and they aren’t the same act.

Will a company find out I used an AI interview copilot?

Sometimes, usually from a human noticing rather than software catching it. CNBC reported that interviewers have learned to watch for eyes wandering to the side or the reflection of other apps during video calls. More employers are also now stating outright, in writing, that any AI use during live interviews breaks the rules, which makes detection less relevant than the policy itself.

Is it cheating to prepare for an interview with AI, even if you don’t use it live?

No. Every policy we could find, including Anthropic’s own published guidance, explicitly welcomes AI use during preparation: drafting practice answers, researching a company, rehearsing a STAR story. The ethical question in this piece is about live, real-time use during the interview itself, not the prep that happens before it.

Lee’s critique of the interview format might even be right. That still isn’t the same claim as “therefore using AI to fake it isn’t cheating.” Those are two separate arguments, and conflating them is how a reasonable complaint about LeetCode turned into three rescinded job offers.