The best evidence of AI-assisted work isn't the text: it's the typing rhythm behind it.
A new study analyzed three datasets: 1,447 recorded writing sessions, 243 programmers' coding sessions, and a 5.1 million-keystroke archive of pre-AI coding used as a human-only baseline. Researchers tracked the timing of individual keystrokes to see whether AI assistance leaves a detectable pattern in both prose and code. It does: AI-inserted text and code arrive in bursts far outside a person's normal typing pace, with the effect roughly three times stronger in code than in writing. But writers and programmers treat that output very differently: 93% of AI-suggested text survived into finished essays, versus just 14% of accepted code suggestions surviving unedited.
That gap matters because most AI detectors today just guess at a finished document, and they're notoriously unreliable and easy to dispute. A classifier built only on keystroke timing correctly flagged simulated wholesale cheating, meaning AI wrote the entire assignment, with near-perfect accuracy, wrongly accusing real students in well under 1% of cases.
It can't yet tell ordinary AI help from unassisted work, and it's only been tested on lab datasets, not a classroom where a false accusation actually ruins someone's semester.