AI Moved the Cheating Into the Live Interview. Small Teams Can Still Read the Real Signal.

The Cheating Moved to the Live Interview

The tools driving this are unusually good at hiding. Apps like Cluely and Interview Coder render a transparent overlay that the candidate can see but screen-sharing software cannot — so on your side of the Zoom or Meet call, nothing looks amiss. The app captures the interviewer’s audio, runs it through a large language model, and surfaces a suggested answer, complete with code, in one to two seconds behind a hotkey. Cluely was founded in 2025 by a Columbia student who had been suspended for building an earlier version that fed AI answers during technical interviews, and it has since raised a $15M round from Andreessen Horowitz. This is not a fringe hack; it is a funded product category.

What makes it corrosive is the exact place it strikes. A structured interview is meant to convert a claim on a CV into evidence: not “I know React” but watching someone reason through a real problem, get stuck, and recover. A real-time overlay collapses that back into a claim. You are no longer watching a person think. You are watching them read.

What the Numbers Actually Say

The scale is easy to underestimate. Fabric, an AI-interview platform, analysed 19,368 live interviews run between July 2025 and January 2026 and flagged roughly 38.5% of candidates for cheating behaviour — a rate it reported had tripled over the final three months of that window. Whatever quibbles you have with any single detection method, a number that large and moving that fast is not noise.

The pattern isn’t uniform, and the shape of it matters. Fabric put the rate at around 48% in software-engineering interviews against roughly 12% in sales, and found junior candidates — zero to five years of experience — cheating at nearly double the rate of senior ones. That tracks with where the tools started (coding screens) and who feels the most pressure (people trying to break in during a brutal entry-level market). The most sobering figure for anyone who screens: Fabric reported that 61% of the flagged candidates still scored above the passing bar and would have advanced to the next round with no detection at all. The signal you think you’re reading in an unstructured call may already be someone else’s autocomplete.

The Detection Arms Race Is a Trap

The instinct is to fight tool with tool: proctoring software, eye-tracking, keystroke analysis, voice-liveness checks, deepfake detection. A whole industry of “interview integrity” monitoring has sprung up to sell exactly this. For a large enterprise running thousands of screens, some of it may be worth the money. For a small team, it is mostly a trap.

It’s a trap on cost, because you become a permanent subscriber to an arms race you can’t win — every detector trains a better evader, and the overlays that got famous are precisely the ones being reworked to dodge the tools that name them. And it’s a trap on trust, which is the one asset a small team actually has. Turning every interview into a surveilled exam tells your strongest, most honest candidates that you assume the worst of them — and those are the ones with other offers, the first to walk. You’d be spending your scarce goodwill to catch the very people you don’t want to hire anyway, while taxing the ones you do.

Structure Beats Surveillance

The better defence is older than the technology, and it happens to be the thing small teams can do best: make the interview hard to fake instead of trying to catch the fakers. A real-time overlay is superb at producing a clean, generic answer to a generic question. It is far weaker the moment you ground the conversation in something specific and follow the thread. Ask candidates to walk through a decision they actually made — then push on the parts that don’t add up. “Why that approach and not the obvious alternative? What broke first? What would you do differently now?” Latency and genericness give the copilot away; a person who did the work has the messy, specific detail no overlay can invent on a two-second delay.

The catch is that this kind of interviewing is hard to do consistently when you’re the only one doing it. Under time pressure, structured interviews quietly drift back into vibes: the same three questions, half-remembered later, scored on a feeling. That inconsistency is the real gap the cheating tools walk through — not a lack of surveillance. It’s the layer we built Kynto to carry. It turns the role into a defined set of criteria before anyone gets on a call, so every interviewer is probing for the same evidence rather than improvising. Its meeting bot joins the interview, produces a speaker-attributed transcript, and afterwards asks each interviewer a structured question set tied to those criteria — then builds the candidate’s score primarily from that human feedback, cross-checked against what was actually said. The point isn’t to police the candidate. It’s to make your own process consistent and evidence-based enough that a canned answer has nowhere to hide. You can see how it fits together at kyntoai.com.

Key Takeaways

• The live interview is no longer a reliable backstop. Invisible AI overlays feed candidates real-time answers, and one analysis of 19,368 interviews flagged about 38.5% for cheating — with 61% of those still clearing the passing bar undetected.
• The surveillance arms race is a bad trade for small teams. Detection tools are an endless subscription and they spend the one asset you have that a multinational doesn’t — the trust of your best candidates.
• Structure is the defence you can actually win with. Role-specific questions, real follow-ups, and scoring against defined criteria expose the generic, delayed answers an overlay produces — without treating everyone like a suspect.

The uncomfortable truth is that AI didn’t break the interview so much as it exposed how loose most interviews already were. A vague, improvised conversation was always easy to game; the overlays just made it obvious. The teams that come out ahead won’t be the ones with the most cameras trained on candidates. They’ll be the ones who decided what “good” means before the call, asked for evidence instead of assertions, and wrote down what they saw. That was better hiring before any of this. Now it’s also the defence.