Invisible AI Overlays Are Quietly Breaking Technical Interviews
Invisible GPU-level AI overlays are undermining technical interviews and screen-share monitoring. Here's what actually still works for engineers.
A new class of tools—Cluely, Interview Coder, Final Round AI—renders LLM-generated answers directly into a candidate's GPU frame buffer, beneath the layer that Zoom, Teams, or Meet capture for screen sharing. The interviewer sees a clean IDE; the candidate sees the same screen plus an invisible AI panel. Those pixels never enter the transmitted video stream, making tab-switch alerts and screen recordings useless as detection methods.
An analysis of 19,368 interviews by Fabric found that 48% of technical candidates triggered cheating flags, and 61% of flagged candidates still scored above the passing threshold and advanced. Despite 64% of companies banning AI in interviews, 80% of candidates use it anyway on take-homes. AI-detection classifiers produce inconsistent, often biased results, and biometric approaches like eye-tracking and facial analysis have proven ineffective or discontinued outright—pushing many firms back toward costly, unscalable in-person interviews.
The most resilient signals turn out to be behavioral and conversational: asking candidates to explain a specific line of code, tracking response-time variance, and relying on system-design or live-debugging formats that can't be answered from a static overlay. For data engineering specifically, framing questions around a real, evolving pipeline failure remains the hardest thing for AI tools to fake.