Somewhere between July 2025 and January 2026, a company called Fabric analyzed over 19,000 live technical interviews and found that 38.5% of candidates triggered signals consistent with AI-assisted cheating. By September, the monthly rate had climbed to 45% and never really came back down. For technical roles specifically, other tracking puts the number even higher — close to half of all candidates. Separately, Karat, a company that runs technical interviews at scale, estimates that roughly 80% of candidates use an LLM during coding assessments even when it’s explicitly against the rules.
This isn’t a rare edge case anymore. For a lot of interviewers, an undetected AI-assisted candidate is closer to the median experience than the exception. And the response from major companies has been telling: Google has reinstated at least one in-person interview round, Amazon now requires candidates to sign a pledge against unauthorized AI use, and Gartner reports that over 70% of recruiting leaders have brought back in-person interviewing specifically to fight this. The remote, purely virtual interview loop that became standard over the last several years is quietly getting rolled back, and this is the reason why.
The tools have gotten genuinely sophisticated. Apps like Interview Coder and Cluely — the latter launching in 2025 under a tagline that literally said “cheat on everything” and pulling 70,000 signups in its first week — listen to an interview in real time, send the question to an LLM, and render the answer invisibly on the candidate’s own screen, positioned specifically to sit below what standard screen-sharing tools capture. A candidate can, in principle, read a fully AI-generated answer off their own screen while appearing to type it themselves, live, on camera.
That sounds close to undetectable, and for a while, functionally, it was. But the data on what actually gives cheating away turns out to be more interesting — and more useful for honest candidates — than the arms race between cheating tools and detection tools.
Across every source tracking this, one detection method keeps coming up as the most reliable, and it’s not a fancy proctoring algorithm. It’s follow-up questions. If a candidate can’t explain, in their own words, why they made a specific choice in code they supposedly just wrote — why this line, why this structure, why this approach instead of an obvious alternative — that’s the tell, every time, regardless of how good the AI-generated answer itself looked. Recruiters who track this closely describe the exact same pattern repeatedly: a flawless submission, code that works on the first try, and then complete silence or vague deflection the moment someone asks “walk me through why you structured it this way.”
This lines up with a second reliable signal: response timing. AI-generated answers tend to arrive with suspiciously uniform latency — the same three to five seconds on nearly every question, regardless of difficulty — because that’s how long it takes a tool to query a model and render a response, and it doesn’t vary the way genuine human thinking time does across a hard problem versus an easy one. Keystroke patterns tell a similar story: AI-assisted typing tends to arrive in clean, confident bursts with almost no backspacing, which isn’t how anyone actually writes code under real pressure, uncertainty, and the normal process of second-guessing yourself mid-line.
Here’s the detail buried in the data that matters most, depending on which side of this you’re on: a meaningful share of candidates who triggered cheating signals still passed the interview and advanced anyway, because detection wasn’t perfect and some interviewers didn’t push hard enough on follow-ups to catch what the behavioral signals were already flagging. That’s genuinely bad news if you’re an honest candidate competing against someone whose Zoom overlay is feeding them answers in real time.
But it’s also exactly why the follow-up question is becoming the actual center of gravity for technical interviews going forward, more than it’s ever been before. Companies that have gotten burned by this are explicitly redesigning their process around it — some banning AI outright and leaning hard on live, conversational rounds where every decision has to be explained and defended in real time; others taking the opposite approach and requiring candidates to use AI openly, then grading specifically on how they reason with it rather than whether they used it at all. Either version converges on the same underlying test: can you actually explain and defend your own thinking, live, under a question you didn’t see coming.
Here’s the reframe that should change how you feel about all of this if you’ve been doing real prep instead of planning to lean on a tool: every single defense companies are building against AI-assisted cheating is, structurally, the exact same skill that’s always separated strong candidates from weak ones — the ability to hold your reasoning together under a “why” you weren’t expecting. This is precisely the gap covered in The Senior Java Interview Depth Guide: Surviving the Follow-Up and in The System Design Interview Library: Defend Your Architecture Under Pushback — both built around a version of the exact test that’s now becoming the industry’s primary defense against AI-assisted fraud.
The candidate who’s genuinely internalized twelve coding patterns instead of memorizing five hundred isolated solutions — the approach covered in 12 Coding Patterns That Cover Most Senior Backend Interviews — is going to sail through a “walk me through why” follow-up, because they actually understand the reasoning, not just the output. The candidate who leaned on an invisible overlay is going to hit that exact same question and have nothing real to say, no matter how clean the code on screen looked thirty seconds earlier.
There’s a harder truth buried in this data too, and it’s worth being honest about it: junior candidates are cheating at roughly twice the rate of senior ones, which suggests a real risk is quietly building in the pipeline — companies unknowingly hiring engineers who can produce AI-assisted output but genuinely struggle with independent debugging, incident response, or architectural reasoning once real, unscripted production pressure shows up, none of which comes with an invisible overlay to lean on. If you’re early-career and you’ve been leaning on these tools to get through interviews, the actual cost isn’t getting caught in the moment — plenty aren’t. It’s landing a role you’re not actually prepared for, and discovering that gap for the first time during your first real incident instead of during a low-stakes practice session where it would’ve been far cheaper to find out.
The interview loop is shifting, visibly and quickly, toward exactly the kind of scrutiny that can’t be faked by a tool feeding you answers in real time. Follow-up depth, live architectural defense, and genuine reasoning under pushback aren’t just good interview skills anymore — they’re becoming the entire point of the exercise, specifically because they’re the one thing an overlay can’t fake for you.
The Senior & Staff Interview Library covers all three tracks — coding, system design, and communication — built specifically around defending your reasoning live, which happens to be the exact skill this entire AI-cheating wave has made more valuable, not less. Companies aren’t just testing whether you can produce a correct answer anymore. They’re testing whether the person in front of them actually understands it. That test was always the real one. It’s just gotten a lot harder to fake your way past it.
38% of Candidates Are Cheating With AI in Interviews Right Now. was originally published in Stackademic on Medium, where people are continuing the conversation by highlighting and responding to this story.