How to check if an essay was written by AI

Every teacher who has set an essay in the last couple of years has probably asked this question. The submission is too tidy, too confident, arrives with none of the usual signs of struggle. But being suspicious and being right are different things, and the gap between them can cause real harm to a student who wrote their work honestly.

The problem with text-based detection

Text-based AI detectors are the first place most teachers look, and they're not without value as a preliminary filter. But their accuracy limitations are well-documented and have caused genuine problems for honest students. The core issue is that these tools flag prose based on statistical patterns - whether word choices look 'surprising' relative to what a language model would produce. Students who write carefully, in formal academic language, or in their second or third language produce text that triggers the same patterns.

Detectors have been repeatedly shown to flag non-native English writers at significantly higher rates than native speakers. A Stanford study tested seven popular detectors and found false positive rates of up to 61% for essays by non-native writers, compared to much lower rates for native speakers doing the same assignment. Several universities have responded by suspending text-based detection entirely.

A false flag on an honest student's essay isn't a minor inconvenience. In formal misconduct proceedings, it can lead to failed grades, official records, and for international students, implications for their visa status. These consequences are serious enough that any detection method needs a higher standard of reliability than current text-based tools can consistently provide.

What a genuine essay process looks like

An essay that a student actually wrote has a recognisable process signature. The session runs long relative to word count: forty-five minutes to an hour for a 500-word piece is fairly typical. Typing happens in bursts with natural pauses - the student stopping to think, rereading a sentence, looking something up. The text grows incrementally, with edits scattered throughout rather than appearing only at the end.

This is what genuine cognitive work looks like. It takes time, it's messy, and the rhythm of the session reflects the rhythm of thinking. Students who are genuinely wrestling with a topic also show more uneven typing patterns - they write faster when ideas flow, slower when they're uncertain.

What AI-assisted shortcuts tend to look like

The process signature of AI-generated work submitted directly - or pasted in after minimal editing - often looks quite different. Common patterns: sessions that last only a few minutes for a submission of several hundred words; a single paste event accounting for most of the word count appearing early in the session; typing that is unusually even and fast, more consistent with transcription than composition; and almost no deletion or revision activity.

These aren't proof of anything by themselves. A student who drafts in a separate document before pasting into an assignment tool will show a large paste event. But when several of these patterns appear together in a submission that also seems inconsistent with previous work, it's reasonable to look more closely.

How to collect process evidence

The most straightforward way to collect process evidence is to set essays through a tool that captures writing behaviour during submission. Some assignment tools record timing data and paste events as standard; others are specifically designed for academic integrity purposes. Learnaway, for instance, captures a timeline of writing events - keystroke timestamps, paste events with sizes, focus changes - without recording the actual characters typed.

Without a dedicated tool, the evidence is harder to gather. You might ask students to submit documents showing revision history, or arrange for supervised writing. These have their own constraints and don't scale easily to regular homework.

Asking the right questions

Once you have some process data, a conversation is almost always more useful than an immediate accusation. 'Tell me how you approached this essay' works better than 'Did you use AI?'. A student who wrote the work can usually walk you through it: where they started, what was difficult, what they changed and why. That kind of process memory is hard to fabricate convincingly.

Specific questions are more revealing than general ones. If the process data shows a large paste at minute three, ask about it: 'I noticed a large block of text appeared early in your session - can you tell me more about that?' This gives the student a chance to explain a completely legitimate workflow before you draw any conclusions.

When the signals add up

Three things together make a stronger case than any one of them alone: anomalous process data; significant inconsistency with the student's previous work quality; and an inability to explain specific choices in the submitted essay. If all three are present, you have reasonable grounds to take it further formally.

Whatever you find, document it before making any formal referral. A timestamped process log is far stronger evidence in misconduct proceedings than a prose score from a text detector. The log is a record of what happened; the detector score is a probabilistic claim about statistical patterns. In any formal context, that distinction matters enormously.