You're thirty seconds into a technical question. The candidate hasn't blinked. Haven't shifted. Their answer comes back polished, accurate, and somehow exactly the right length. Then you ask a follow-up. Dead air. Long pause. The quality drops.
Something feels off.
Remote interviews have made it easier to spot cheating and harder at the same time. Studies show 35% of candidates display signs of cheating during remote technical interviews, and dedicated tools account for 45% of detected cases. The real question is whether you can catch it-and more importantly, whether you can document it when you need to.
What manual red flags actually tell you
Response timing tells a story. When a candidate uses an AI overlay or audio loopback, the answers come back in a consistent 1-2 second window. Every answer. No variation. That consistency is the red flag, not the speed itself.
Eye movement reveals what they're reading. When someone is reading an answer from an off-screen source, their eyes stay locked in one direction. Watch for the candidate whose gaze never wavers from the lower left corner of their screen.
Answer quality inconsistency cuts both ways. You ask a highly technical question-the answer is immaculate. Then you ask: "Why did you choose that approach?" Silence. The follow-up reveals the first answer wasn't actually understood, just retrieved.
Structure and polish matter when it's excessive. AI language models tend to produce answers that are smooth, polished, and a little too perfect. When every answer follows the same pattern (introduction, three main points, summary), it's worth noting.
Why manual detection fails most of the time
An overlay application running at the GPU level operates beneath the screen-sharing layer. You literally cannot see it. Audio routing is even more invisible. A candidate can pipe ChatGPT's audio output directly into their microphone line using software that costs nothing.
This is why 45% of detected cheating requires dedicated tools. Manual observation catches maybe the obvious cases. Everything else needs to be surfaced by the technology.
How to actually detect AI assistance
Device integrity monitoring checks what's running on the candidate's computer. Are there overlay applications installed? Are there known prompt-injection tools in memory?
Browser behavior analysis reveals what tabs the candidate has open, whether they're switching between windows during answers, and the pattern of their keyboard and mouse activity.
Audio routing analysis checks how the microphone input is configured. Normal speech comes from the candidate's microphone. AI audio loopback produces a different signature.
Detection alone doesn't protect you
Catching cheating matters. Proving you caught it matters more. Detection generates data-process lists, audio characteristics, device state, behavior patterns. That data lives in a report. When you need to justify a hiring decision, you have evidence. The difference between "we think something might have happened" and "here's exactly what we detected" is the difference between a hiring decision that holds up and one that collapses under scrutiny.
ScreenComply automates AI detection and generates audit-ready documentation in a single platform. Instead of guessing whether a candidate used AI, you get technical evidence tied to each interview.
For take-home case studies and modeling tests, see How to tell if a take-home case study was done with AI.