How to detect AI writing in student homework
The most defensible way to detect AI writing in homework is to look at how the work was entered rather than how it reads. Behavioural signals (a large paste, an unnaturally uniform typing rhythm, a long essay finished in minutes, or text appearing with no keystrokes) are concrete and explainable, whereas text-based AI detectors frequently misclassify human writing, especially from non-native speakers.
Why text-based detectors fall short
Tools that estimate AI use from prose style rely on statistical quirks that disappear as models improve and that unfairly flag simpler or non-native writing. Universities including Vanderbilt disabled such detectors after high false-positive rates.
Behavioural signals that actually help
Watch for: a high proportion of text arriving via paste; very low variation in keystroke timing (a sign of transcribing); submission far faster than plausible for the word count; and large blocks appearing late in the session with no typing before them. Any one alone is weak, but multiple together are meaningful.
Treat signals as a prompt for a conversation
No signal proves misconduct. A student may paste from their own notes or use assistive technology. Use signals to decide which submissions deserve a closer look, then talk to the student.
FAQ
- Can you prove a student used AI?
- No tool can prove it. Behavioural signals give you defensible evidence to inform a professional judgement and a conversation, not a verdict.
- What's the single strongest signal?
- A large share of the final text arriving in one or two paste events, corroborated by an implausibly short writing time.
Want to see this on real writing?