Take-home coding tests were the gold standard of engineering hiring for a decade. Then GPT-4 and Copilot happened. Now anyone can ship a passable solution to a typical take-home in an afternoon — even if they can't actually code.
Anti-AI detection plus reframing the test format is how teams keep the signal honest. Pure detection can't be the whole answer — false positives are too costly. But layered with format changes, you can keep take-homes useful.
First, behavioral signals: typing cadence, edit patterns, paste events, time-on-task distribution. AI-assisted submissions look different from human-written ones. Detection accuracy on our benchmark sits at 94%, with false-positive rate <2%.
Second, format change: shift from 'implement X' to 'review and improve this codebase'. AI is bad at debugging unfamiliar code, especially with subtle architectural issues. Real engineers can do it; LLM-assisted candidates will struggle.
Third, follow-up live walkthrough: 30-minute call where the candidate explains their solution. Anyone who actually wrote it can talk fluently about tradeoffs; LLM-assisted candidates can't.
