Somewhere between "AI will fix hiring" and "candidates will use AI to fake their way through hiring" is where the industry actually landed in 2026. Both are true at once, and most teams evaluating AI interview tools are only prepared for the first one.
The numbers are not subtle. Research from Fabric analyzing 19,368 interviews between July 2025 and January 2026 found that 38.5% triggered some kind of cheating flag — and that rate jumped roughly 3x in a single quarter, from 9% in July to 45% by September, and has stayed elevated since. Separately, surveys keep landing near the same figure: 83% of candidates say they'd use AI assistance in an interview if they thought they could get away with it.
So the honest starting point for any AI interview product isn't "how do we detect cheating." It's "how do we design an interview where cheating stops being useful in the first place."
The tooling has gotten specific. Tools like Cluely, Interview Coder, Yoodli, and Pickle create transparent overlays that feed scripted or AI-generated answers directly onto a candidate's screen — built specifically to be invisible during screen sharing. Some candidates run a second monitor or a phone below the camera frame. Others open a second browser tab with ChatGPT and read from it while maintaining eye contact with the camera.
What almost all of these methods have in common: they work best when the interview is static. A known question, unlimited think time, and no adaptive follow-up is exactly the environment scripted AI assistance thrives in.
A wave of anti-cheating products has shown up to answer this — eye-gaze tracking, screen and tab-switch monitoring, keystroke dynamics, response-timing analysis, even detecting the subtle shift in speech pattern when someone starts reading a generated answer aloud (vocabulary flattens, filler words disappear, sentence structure regularizes).
These are useful signals. They're also an arms race — every detection method spawns a workaround, and legitimate candidates get flagged by false positives (a nervous pause reads the same as a suspicious one to a timing model). Proctoring software can tell you that something looks off. It's much weaker at telling you whether the candidate can actually do the job.
Detecting cheating and evaluating competence are different problems. Solving only the first one still leaves you with a bad hire — just one who didn't get flagged.
Building Ray, we treated this as a design problem before a detection problem. A live, voice-based, adaptive interview structurally removes most of the easy attack surface that static formats have:
None of this makes cheating impossible. Nothing does, and any vendor claiming otherwise is overselling. But it changes the economics for a candidate — a static script gets you through a static interview, and adaptive follow-up is expensive to fake in real time, especially across a full 30-45 minute session covering screening, DSA, and system design.
Ask any vendor two questions: does the interview adapt in real time based on what the candidate just said, and does the final report show how a candidate reasoned through pressure, or just a score. If the answer to either is no, the tool is probably better at running interviews at scale than at resisting the exact failure mode candidates are now actively engineering for.
The goal was never to build an uncheatable interview. It was to build one where the fastest way to pass is still to actually know the material.
Ray is currently onboarding early access teams. Join the waitlist to be among the first to try it.
Join the Diyam AI Waitlist →