How to spot AI writing in student work
To spot AI writing in student work, look at how the text was entered, not how it reads. The clearest indicators are a large paste event delivering 60–90% of the final word count, an inter-keystroke variance below 150ms (consistent with transcribing AI output rather than composing), and a total session time implausibly short for the length - a 1,000-word essay completed in under 8 minutes is a meaningful outlier. Two or more of these behavioural signals together are more reliable and fairer than any prose-style AI checker.
Why prose-style checkers struggle to spot AI writing
Tools that read prose to estimate AI origin chase statistical patterns that improve AI models quickly erode. They also penalise students with simple or non-native writing. Several universities have limited text-only AI checkers in disciplinary proceedings because false-positive rates were too high to sustain.
The behavioural signals that reveal AI-written work
The most informative evidence is in the writing timeline: a single paste event covering 60–90% of the final word count; inter-keystroke variance below 150ms, suggesting transcription rather than original composition; a total session under 8 minutes for a 1,000-word assignment; and little or no editing activity after large text insertions. Any one signal is weak in isolation; two or more together are meaningful.
Acting on signals: conversation, not verdict
No behavioural signal proves AI use. A student may have pasted from their own notes, used assistive technology, or written the work in another app first. Use flagged signals to decide which submissions warrant a closer look, then speak with the student. Document the specific data points - for example, '74% of text arrived in one paste; session lasted 6 minutes for an 800-word essay' - to support a professional, evidence-based conversation.
FAQ
- What's the single strongest signal for spotting AI writing?
- A large share of the final text arriving in one paste event, corroborated by a submission time implausibly short for the word count. For a 1,000-word essay, a total writing session under 8 minutes is a meaningful outlier.
- Can a student disguise AI-written work from behavioural detection?
- A student who manually re-types AI output can suppress paste signals, but unusually uniform inter-keystroke timing - consistent gaps below 150ms - can still hint at transcription rather than original composition. Behavioural signals are harder to fake than prose style.
- Does behavioural detection disadvantage non-native English speakers?
- No. Because it ignores writing style entirely, behavioural detection doesn't recreate the ESL-bias problem that led universities to restrict text-based AI checkers. A non-native writer who types their own work won't trigger paste or timing signals regardless of their prose.
Want to see this on real writing?