How Interview AI Detection Works, and Where It Fails

On June 20, 2025, a company whose whole pitch is helping people cheat closed a $15 million round led by Andreessen Horowitz. Cluely, run by 21-year-old Roy Lee, the Columbia student suspended a year earlier over a tool called Interview Coder, reportedly landed at around a $120 million valuation, TechCrunch reported. That was roughly two months after its $5.3 million seed. Money is moving fast on the offense side of interview cheating. Most candidates have no idea how interview AI detection works on the defense side.

That defense side is what this post is about, and it’s more mechanical than philosophical. If you’re a candidate weighing an AI copilot, you don’t need a lecture on ethics. You need to know how the detection actually functions, method by method, and which parts are real technology versus stuff people repeat to each other in Discord servers. So this walks the plumbing: gaze monitors, video-interview face analysis, keystroke telemetry, and the oldest detector of all, a suspicious human on the other end of the call. Every mechanism below links to a primary source, so you can check it without taking my word for anything.

How Interview AI Detection Works, in One Paragraph

Interview AI detection works in three overlapping layers. Software watches your webcam and screen for physical tells like gaze direction, extra faces, or a second monitor. Platform telemetry reads your typing rhythm, paste events, and how long you take to answer. And a live interviewer, increasingly in the same room, listens for the pause-then-flawless-answer pattern. No single layer decides anything on its own; most flags come from two of them lining up.

That structure is the whole point, and it’s where most public arguments go wrong. People fixate on one layer, usually the screen, and ignore the other two. A tool can be perfectly hidden from a screen share and still get someone flagged on cadence, or caught cold by a follow-up question a copilot can’t answer for them. The rest of this walks each layer and what the evidence says it can and can’t do.

The Eye-Tracking Myth: What Gaze Detection Actually Measures

Start with the fear you hear most, that proctoring software tracks your eyeballs. It mostly doesn’t. Proctorio, one of the larger remote-proctoring vendors, spells this out in its own FAQ: it uses gaze detection “to see if a test-taker is looking away from the screen for an extended period of time,” and face detection “to detect the presence of one or more human faces or if the test-taker has left the exam.” That’s orientation and presence, not pupil tracking.

The distinction is bigger than it sounds. Gaze detection can tell that your head swung toward a second screen for eight seconds. It can’t tell what you read there, and Proctorio’s FAQ is explicit that it doesn’t use facial recognition and can’t uniquely identify a face. So the mental picture of an AI following your irises across a monitor, ranking every micro-glance, is closer to campus folklore than to what these tools ship. What they actually flag is coarse: you left the frame, a second person appeared, your face pointed away from the screen too long.

The classic tell, eyes darting to a second monitor, is real. But it’s a blunt instrument that catches the obvious cases and misses the careful ones.

What Video Interviews Tried to Read From Your Face, and Why the Biggest Vendor Mostly Quit

There was a stretch where the industry believed a camera could read competence off your face. HireVue, the video-interview company used across a big share of large-employer hiring, built AI that analyzed facial muscle movements during recorded interviews, things like furrowing your brow or smiling. In January 2021 it announced it had stopped. Both SHRM’s report and Fortune’s coverage date the actual change to March 2020.

The reason is the part worth sitting with. HireVue’s own chief data scientist told Fortune that nonverbal data contributed roughly 0.25% to the predictive power of the assessment in most roles, and about 4% for customer-facing jobs. For a feature that made candidates dread being scored on their expressions, that’s almost nothing. The face-reading was oversold, and the company’s own numbers said so.

There was outside pressure too. The Electronic Privacy Information Center had filed an FTC complaint in 2019 calling the practice unfair and deceptive, and an independent audit by O’Neil Risk Consulting flagged bias problems, including that the system handled different accents inconsistently and struggled with minority candidates who gave shorter answers. Facial analysis in interviews didn’t get beaten by a clever candidate. It got dropped because it barely worked and generated more risk than value.

I think that history is the most useful single thing a candidate can know about AI interview detection technology. The flashiest layer, the camera claiming to read your face, is the one the industry already walked away from. What replaced it is quieter, and harder to fool.

How Do Companies Detect AI in Interviews When Nothing Shows on Your Face?

When face-reading fails, companies fall back on two things that work better: telemetry from the assessment platform, and a human. Coding platforms log keystroke cadence, paste events, and how often you leave and re-enter the test window. And a rising share of employers have moved interviews back into a room, where a trained interviewer is the detector.

The in-person swing is the clearest signal of where detection is heading. Computerworld reported in 2025 that Google, Cisco, and McKinsey were all reintroducing face-to-face interviews specifically to counter AI-assisted cheating and fake candidates. Google banned AI tools during virtual interviews and started bringing engineers on-site earlier in the process. Cisco’s VP of Global Talent Acquisition, Scott McGuckin, put it plainly: “Remote work and advancements in AI have made it easier than ever for fake candidates to infiltrate the hiring process.” A McKinsey spokesperson framed in-person rounds as a way to judge “human qualities that can’t be automated,” like judgment and empathy.

The scale is real, not anecdotal. Citing Gartner, the same report noted that 72.4% of recruiting leaders now conduct in-person interviews to combat fraud. No overlay, no content-protection flag, no screen trick touches that number. A live human across the table is a detection method with zero software dependency, and it’s the one companies are quietly betting on.

The Three Places You Actually Get Caught

Stripped down, every method above sorts into one of three layers. This framing is mine, built from the sources in this post, and I’d hold it loosely. But it has matched every real case I’ve read about so far.

My opinion, and it’s arguable: candidates over-index on the first row and under-index on the third. The camera layer is the most feared and the weakest, full of blunt flags and documented blind spots. The human layer is the least discussed and the hardest to beat, because a follow-up question doesn’t care what’s on your screen. It cares whether you actually understand the answer you just gave.

Where a Tool Like LastRound Fits in This

Plenty of products now build for the offense side of this, from Cluely down to smaller copilots. LastRound is one of them, with an AI Interview Copilot. I’m not going to tell you whether it beats any specific detector named above, because that’s not a claim anyone can honestly make in the abstract, and detection shifts with every platform update. What’s worth quoting instead is a line from LastRound’s own help center, the kind of caution most vendors would cut: it says the desktop app “has not been tested on all coding platforms” and tells users to “verify that your interview platform permits the use of external assistance tools before starting your session.” That’s the tool’s own documentation telling you to read the rules first, which is sound advice no matter which product you’re looking at.

FAQ

Can proctoring software track my eye movements?

Mostly no. Major proctoring tools like Proctorio use gaze detection, which checks whether your face is turned away from the screen for an extended period, not fine-grained eye or pupil tracking. Proctorio’s own FAQ says it doesn’t use facial recognition and can’t uniquely identify a face. It flags orientation, extra people, and leaving the frame, not where your eyes point.

How do companies detect AI in interviews in 2026?

Through a stack, not a single tool. Webcam and screen monitoring catch physical tells, assessment platforms read typing and paste telemetry, and a rising share of employers have moved rounds back in person. Gartner found that 72.4% of recruiting leaders now run in-person interviews to combat hiring fraud, according to Computerworld. The trend line points away from clever software and toward a human in the room.

So the detection stack in 2026 is less a wall than a set of overlapping nets, each with holes the others are meant to cover. The camera net is the weakest and the most feared. The human net is the strongest and the least talked about. If you’re deciding whether an AI copilot is worth the risk, the honest place to start isn’t “can it hide from a screen share.” It’s “who is actually watching, and would I make the same call if they’d told me in advance?”