Fabric AI tracked 19,368 live interviews between July 2025 and January 2026 and flagged 38.5% of candidates for AI assistance. In software engineering the rate was 48%. In sales, 12%. The July rate was 9%. By September it was 45%.

The tools behind the curve are commercial products with pricing pages. Cluely and Interview Coder run invisible overlays that listen to the interviewer and print an answer next to the camera, and they claim over a million users between them. Detection vendors answered with gaze tracking and response-timing analysis. Candidates answered the detectors with earpieces and deepfakes. When Fabric asked, 83% of candidates said they would use live AI help if they were confident nobody could tell.

One number in that report matters more than the 48%. Of the candidates flagged for cheating, 61% scored above the passing threshold and would have advanced anyway. Read that from the candidate's side: the fastest way to pass your interview is to get an answer out of a model and deliver it with a straight face. Your loop selects for that.

The uncomfortable part is that getting a working answer out of a model is the job now. A candidate who can steer a model to a passing answer under pressure has a skill your team uses daily. The overlay hides the part you need to see: did they check the answer, do they know why it holds, would they catch the case where the model lies.

So stop building interviews an overlay can beat. If the test is "produce this answer," software produces it. Hand candidates a real codebase, put the model on the table, and watch the direction.

That is a Crackd screen. AI is allowed and logged. Candidates work a real codebase for 90 minutes, and the grade comes from the judgment trail: the prompts they wrote, the output they rejected, the bug they caught that the model missed. Cluely can't help, because using the model well is the test. The 48% are your applicant pool. Build the interview that tells them apart.