A Dropout Raised $15M to Cheat Your Interview. He Is Half Right.

In 2025, a 21-year-old Columbia student named Chungin "Roy" Lee built a small piece of software called Interview Coder. It did one thing: it put a hidden window on your screen during a remote technical interview — invisible on a screen share — that watched the coding question and fed you the answer. He used it to land a software engineering internship at Amazon.

An Amazon interviewer figured out what had happened and reported it to Columbia. The message, by Lee's own account, was blunt: expel this kid, or we stop hiring from your school. Columbia hauled him into disciplinary hearings and suspended him.

Most people would treat that as a cautionary ending. Lee treated it as a launch. He rebranded the tool as Cluely, gave it a one-line mission — "cheat on everything" — and turned the suspension into a marketing campaign. By April 2025 he'd raised $5.3 million in seed funding. By June, Andreessen Horowitz led a $15 million round at a reported valuation near $120 million. The pitch to investors wasn't subtle. It was that the ability to invisibly feed a person the right answer, in any conversation, was a product the whole world would eventually want.

You can be appalled by this. A lot of people are. But the appalled reaction skips the part that should actually worry anyone who hires: the tool worked, it works at scale now, and the reason it works is not that Cluely is clever. It's that the interview was a teleprompter-shaped hole the whole time. Someone finally built the teleprompter.

"Cheat on everything" is not a fringe product anymore

It's tempting to file Cluely under viral stunt — a provocateur founder, a deliberately outrageous tagline, a16z chasing attention. The branding is a stunt. The behavior underneath it is not.

Cluely, Interview Coder, Final Round AI, and a growing field of imitators all do the same mechanical thing: an off-screen overlay transcribes the interviewer's question in real time and generates a polished answer for the candidate to read aloud, with nothing visible on the shared screen. It works for coding challenges, behavioral questions, system design, sales role-plays — anything conducted through a webcam and a screen share, which after five years of remote hiring is most things.

The adoption numbers have stopped looking like a fringe. One analysis of 19,368 technical interviews conducted between mid-2025 and early 2026 found that 38.5% showed signs of AI assistance; in purely technical roles it ran closer to 48%. Across a broader candidate pool, the rate roughly tripled inside a single year — from around 15% in June 2025 to about 35% by December. In one survey of more than 3,600 professionals, one in five admitted to secretly using AI during an interview, and 55% said it had simply become the new norm.

The detail that should end any comfort that this is low-skill desperation: 61% of candidates who used these tools cleared the interview's passing threshold. The cheat isn't a Hail Mary. It's a reliable way to pass. And junior candidates — the ones with the least to show and the most to gain — use it at roughly double the rate of seniors, which means the tool is most concentrated exactly where the signal was already thinnest.

We've written before about candidates who aren't real — the identity fraud, the deepfakes, the manufactured profiles. This is the opposite problem, and in some ways the more unsettling one. These are real people, in real interviews, who genuinely applied. They're just fabricating, live, the one thing the interview exists to measure: their unaided ability to do the work.

The panic: drag everyone back into a room

Faced with a backstop that suddenly doesn't hold, the industry has reached for the most intuitive fix available — remove the screen the cheat lives behind.

The retreat to in-person interviewing is now a real trend with real names attached. Through 2025, Google, Cisco, and McKinsey all reinstated mandatory in-person rounds for at least some roles specifically to counter AI-assisted cheating. Amazon began requiring candidates to formally acknowledge they won't use unauthorized AI tools. Anthropic put an explicit ban on AI use during its application process. By the time Gartner surveyed recruiting leaders, 72.4% said they were conducting interviews in person to fight fraud — a striking reversal for an industry that spent half a decade congratulating itself on going remote.

On its face this looks like rigour returning. Pull the candidate into a room, take away the second monitor, watch their hands. Problem solved.

It isn't, for two reasons, and they compound.

The first is cost and reach. In-person interviewing is expensive, slow, and geographically narrowing. It quietly re-excludes the remote candidate, the caregiver, the person two time zones away, the applicant who can't take a day off and fly across the country on spec. To plug a fraud hole, you shrink the funnel back to people who can show up — which has nothing to do with who can do the job, and a lot to do with who's already privileged enough to be nearby.

The second is that it doesn't even close the hole. Deepfake video and audio have advanced to the point where a fabricated candidate can hold a live conversation; a CBS study found half of businesses had already encountered AI-driven deepfake fraud of some kind. The proctored room buys a few months. Then the cheat moves a layer up, the detector moves a layer up to chase it, and you're back where you started, having spent enormous effort to relocate the problem rather than solve it.

The uncomfortable thing Cluely is half right about

Here is the part the outrage skips, and it's the part that matters.

Strip away the obnoxious branding and Roy Lee's actual provocation is a real question: if the job is done with AI sitting right there — if every engineer, analyst, and marketer at your company will use a model all day to do the work — then what, exactly, is a live interview that bans AI measuring? It's measuring the candidate's ability to perform the job under conditions the job no longer has. It's testing them on a fictional version of the work.

That argument is too convenient when it comes from a man selling the cheat. But it's not wrong, and that's what makes it dangerous to ignore. The interview was never a clean measure of ability. It was always a measure of interview performance — how well someone codes a contrived puzzle on a clock, recites a STAR-format story they rehearsed, stays composed under an artificial spotlight. Plenty of excellent practitioners interview badly. Plenty of mediocre ones interview beautifully. The gap between "good at the interview" and "good at the job" has been the open secret of hiring for decades.

AI didn't create that gap. It industrialized it. Cluely is what happens when you take a test that was already measuring the wrong thing — performance under artificial constraint — and hand everyone a tool that aces the artificial constraint. The reason the cheat is so effective so fast is that the interview was never that hard to fake in the first place. It rewarded smoothness over substance long before there was a window feeding people the smooth version.

So the in-person retreat isn't just costly and leaky. It's a doubling-down on the original error: it spends real money to protect the integrity of a test whose integrity was the problem. You can perfectly proctor a measurement and still be measuring the wrong thing. A clean read of a bad instrument is still a bad read.

Every defense escalates in the wrong direction

This is the same trap automated hiring keeps walking into, and it has a predictable shape.

You add a defense — a proctor, an AI-answer detector, a return to in-person, a signed pledge. The cheaters, who are motivated, technical, and treat every gate as a puzzle, route around it; that's literally Cluely's product roadmap. Roy Lee, asked about detection tools, said he isn't sweating them — because the economics favor the cheat. The honest candidate, meanwhile, experiences your new defense as suspicion: the surveillance, the accusation, the assumption that they're guilty until a detector clears them. AI-answer detectors are notoriously trigger-happy, and a false positive doesn't just lose you a hire — it brands a real person a fraud for the crime of being articulate.

So the gate sorts in exactly the wrong direction. The practiced cheater, indifferent to the friction, pushes through. The strong honest candidate, insulted by being treated as a suspect, drops out — and strong candidates have other offers, so they're the first to walk. You don't clean the funnel. You change who survives it, tilting it toward whoever is most willing to be surveilled and most willing to game the test. Your detection dashboards light up green and the actual quality of who you hire quietly erodes, because the metric improved while the measurement got worse.

You cannot out-proctor a teleprompter. There is always another monitor, another phone, another model, another layer. An arms race you're structurally destined to lose is not a strategy. It's a budget line that grows forever and buys less every year.

What a hidden window can't whisper

If the live, unassisted performance can be faked in real time — and it can, and increasingly is — then the only signal worth building on is the one a teleprompter can't generate.

An AI window can hand you the answer to a coding question. It can write you a flawless story about a time you handled conflict. What it cannot do is invent years of work you didn't do. It can't fabricate the arc of how you actually moved through a career — the problems you owned and the order you owned them in, the projects that shipped and the outcomes that cross-reference against a real world, the way a genuine trajectory connects into a coherent line instead of a stack of rehearsed moments. The interview is a 45-minute snapshot, and any snapshot can be staged. A trajectory is a record you had to have actually lived, day after day, for years. There is no overlay for that.

This is the whole point. The thing that survives a world of real-time cheating is not a harder test or a better proctor. It's a different object of evaluation entirely — the pattern of a real career rather than the performance of a single conversation. You evaluate what someone has demonstrably done over time, how it fits together, what it predicts about the work in front of them. You treat the interview as one small, corroborating data point in a much larger picture, not the picture itself — because the moment the interview is the picture, you've built your entire hiring decision on the one surface a $120 million company exists to forge.

Cluely is genuinely right that an interview testing AI-free ability is testing a fiction. It is genuinely wrong that the answer is to help everyone cheat the fiction. The answer is to stop staking the decision on a performance that was always game-able and start staking it on a trajectory that isn't. A teleprompter can change what a candidate says in the room. It can't change what they actually did before they walked in.

The interview now has a teleprompter. The companies that come out of this well won't be the ones who spent the most trying to confiscate it. They'll be the ones who realized they were never supposed to be hiring the performance in the first place.

AgentR evaluates candidates on the trajectory of a real career — the work actually done, the way it fits together, the pattern that predicts the job — rather than a performance in a single interview that AI can now script in real time. You can't out-proctor a teleprompter, but you can evaluate the one thing it can't fake. If your hiring decision still rests on a 45-minute show, you're measuring the most cheatable surface you have. Let's talk.