I've been on hiring panels where we spent 45 minutes convinced a candidate was sharp. Articulate answers. Clean code. Solid reasoning. Then in the debrief, someone pulled up the recording and timed the responses. Every answer: 4 seconds. Easy question, hard question, curveball follow-up. 4 seconds flat. Humans don't think like that. Humans stumble on hard problems and breeze through easy ones. This person's cadence was perfectly uniform. I'd love to tell you I spotted it in real time. I didn't. Nobody on the panel did.
That was 8 months ago. The tools have gotten significantly better since.
The Overlay That Broke Data Engineering Hiring
There's a new category of AI cheating tools that most interviewers don't know exist. Cluely, Interview Coder, Final Round AI; these aren't browser tabs a candidate Alt-Tabs to when you're not looking. They're invisible GPU overlays that use DirectX (Windows) and Metal (macOS) hooks to render LLM-generated answers directly on the candidate's monitor, beneath the capture layer that Zoom, Teams, and Google Meet use for screen sharing.
In plain English: when you screen-share on Zoom, the application captures pixels from a specific layer of the OS graphics pipeline. These overlay tools render their content below that layer, in the GPU's local frame buffer. The interviewer sees a clean IDE. The candidate sees the IDE plus a floating panel with AI-generated code and explanations. Those pixels literally don't exist in the video stream that gets encoded and transmitted.
This isn't a proof of concept from a security researcher. Cluely pulled 70,000 signups in its first week. It's a consumer product with standard SaaS pricing: free tier at 5 responses per day, $20/month for Pro, $75/month for the "undetectability" tier that does the GPU-level rendering. Cheating on your data engineering interview now costs less than a monthly gym membership.
The overlays don't trigger tab-switch alerts. They don't appear in keystroke logs. They don't show up in screen recordings. The proctoring vendors claiming 85-95% detection effectiveness? They're scanning for browser tab switches and copy-paste events. They're watching the wrong layer entirely. 6 new overlay tools emerged in 2025 alone, plus at least 3 open-source clones.
Screen-share monitoring is security theater. The pixels the interviewer sees are not the pixels the candidate sees.
Half Your Tech Candidates Are Cheating. Most of Them Pass.
Fabric analyzed 19,368 interviews between July 2025 and January 2026 across 50+ companies. The numbers are bleak.
38.5% of all candidates triggered cheating flags. For technical roles, that number hit 48%. Sales roles? 12%. The roles where the stakes are highest and the questions are most googlable are the roles getting gamed hardest. This inverts the assumption that technical interviews are somehow more honest by nature. They're not. They're just more automatable.
The acceleration is what gets me. Cheating went from 9% in July 2025 to 45% by September. 3x in 3 months. Then it plateaued through January; not because people stopped, but because it saturated. When nearly half your candidate pool is using AI assistance, you don't have a cheating problem. You have a broken measurement system.
The part that should make every hiring manager stop and think: 61% of flagged cheaters scored above the passing threshold and advanced in the pipeline. These aren't marginal candidates scraping by. They're clearing the bar comfortably, because the AI is genuinely good at answering the questions we ask. 61% is not a detection gap. It's an action gap. Companies flag candidates and hire them anyway because the score looked fine.
Junior candidates (0 to 5 years of experience) cheat at nearly double the rate of seniors. The people with the least context to evaluate whether an AI-generated answer is even correct are the ones relying on it most. It's a desperation play: they know AI fills knowledge gaps faster than grinding prep, and they're probably right. The incentive structure is broken.
Banning AI Doesn't Fix the Architecture
64% of companies ban AI in interviews. 80% of candidates use it anyway on take-homes. That's not a policy failure; that's a policy being laughed at.
The detection layer makes it worse. AI detection tools produce essentially random results: the same essay scored 4%, 91%, 12%, 67%, and 38% across 5 different detectors. OpenAI killed its own classifier in July 2023 because it was only 26% accurate. Stanford found a 61.2% false-positive rate on essays from non-native English speakers versus 5.1% for native speakers. You're not catching cheaters. You're penalizing people who learned English as a second language.
The industry response has been a stampede back to in-person interviews. They jumped from 5% to 30% in a single year. Google and McKinsey mandated in-person rounds in mid-2025, explicitly citing AI fraud. 72% of recruiting leaders say fraud prevention is the driver.
The problem: in-person doesn't scale. It's expensive, exclusionary, and locks out remote candidates, international candidates, and anyone who can't fly to your office for a day. The least scalable option is the one everyone's reaching for. That's not a fix; it's a retreat to 2019.
Take-homes are dead in their current form. Anthropic's own engineering team found that Claude Opus "matched top candidates" on take-home assessments and there was "no longer a way to distinguish between the output of top candidates and the most capable model." When the company building the AI tells you their model passes your take-home, believe them.
Eye tracking? Gaze monitoring? Also crumbling. NVIDIA Maxine synthesizes natural gaze. Candidates keep their eyes near the camera while reading overlays at the screen periphery. HireVue discontinued facial analysis in 2021 after finding facial expression data contributed less than 0.25% to performance predictions. The biosurveillance approach is simultaneously invasive and useless.
You can ban AI in interviews. You can't ban the architecture it runs on. The enforcement tools are more invasive than the cheating, and they still don't work.
The Interview That Survives
The strongest signal isn't software. It's one follow-up question.
AI generates polished, multi-layered code. But when you ask a candidate to explain line 7, or walk through why they chose that join strategy, or refactor their solution with a new constraint, the gap between "read this off a screen" and "actually understand it" becomes obvious in seconds. If you can't explain line 7, you didn't write line 7.
Consistent response timing is the second strongest behavioral fingerprint. Humans pause longer on hard questions. They stammer, correct themselves, say "let me think" with variable duration. AI overlay tools show uniform 3 to 5 second latency regardless of difficulty, because the pipeline is always: audio capture, transcription, LLM inference, render. That flatness is detectable. But only about 30% of interviewers actively monitor response timing. The other 70% miss it entirely.
System design remains the most AI-resistant format because it's fundamentally discursive. You can't read a system design answer off an overlay, because system design isn't a question with a fixed answer. It's a conversation that shifts when the interviewer changes constraints, pushes back on choices, and asks "what happens when this fails?" No overlay tool fakes that in real time. Not yet.
For data engineering interviews specifically, the path forward is clearer than for most roles. Pipeline architecture, data modeling exercises, debugging scenarios. "Here's a pipeline that silently dropped 2M rows last Tuesday. Walk me through how you'd find the problem." That question has no clean overlay answer because the solution depends on follow-ups about the specific system, the specific data, the specific failure mode. The actual job is debugging, not building; might as well test for it.
This is also why concepts matter more than tools, and always have. If your interview tests whether someone can write a Spark transformation, an overlay solves it in 4 seconds. If it tests whether someone understands why a pipeline breaks when upstream schema changes violate a downstream join contract, you're testing something AI can't fake. Concepts transfer across tools; syntax doesn't, and that gap is exactly why we built databricks interview prep with datadriven around architecture walkthroughs and data modeling, not the kind of timed problems an overlay solves in 4 seconds.
The career implications cut both directions. If you're a candidate who genuinely knows your stuff, push for system design and pair-programming rounds. Those formats favor you. If a company's entire loop is timed LeetCode and async take-homes, their signal is already compromised, and your real expertise is competing against someone paying $75/month for invisible help.
The companies that adapt their process will hire better. The ones clinging to LeetCode mediums and take-homes will keep onboarding engineers who can't debug a broken DAG in their first week, then blame the candidate instead of the process.
What's the most egregious cheating you've seen in an interview, on either side of the table?
Top comments (0)