“I don't try to catch AI use by eye anymore. The tells change too fast, and the real problem is that most coding interviews are set up to be cheatable in the first place.”
“We caught a candidate this way: the first interview was polished, but the systems design round fell apart, they couldn't draw a diagram or defend a single choice. The tells were long pauses, generic phrasing, and answers that felt Googled. I now tell candidates upfront: AI is encouraged on the job, prohibited in the interview, because I need to see what you actually know versus what you can regurgitate.”
“I stopped trying to police AI in interviews because the real signal is whether they can catch AI-generated mistakes under pressure. I now let candidates use AI and watch how they explain prompts, iterate, and own the output.”
“I stopped trying to catch candidates using AI and started designing interviews where using it badly is self-revealing. The dead giveaway is silence: someone pastes code they cannot walk through line by line, or they accept an AI suggestion that breaks an edge case they should have spotted. In the interviews I ran, the strongest negative signal was solution blindness — the code is there, the understanding is not.”
“I built an AI interviewer at Eightfold that had to detect exactly this, so I think about it less as catching cheating and more as designing the loop so AI help doesn't matter.”
“The strongest signal is not what they type, it is how they behave when you push them off script. Ask why they chose a variable name, or change one requirement mid-problem; someone using Copilot will stall or produce code they cannot explain.”
“I spent months building Anti-Soy to spot AI dependency in GitHub profiles, and the real tell in a live interview is the same gap: they can produce code but cannot explain why they chose that structure or what they would change if the constraints shifted.”
“I stopped trying to police AI in interviews the day I realized the real problem was the interview itself. A candidate who can pass your live coding screen with a hidden prompt is just telling you the screen was testing performance under questioning, not the actual work.”