How to tell if a student used AI to write an essay

You're reading an essay and something's off. It's clean — maybe too clean. The vocabulary doesn't sound like the kid who sits in the third row. You want to know whether it's actually theirs. Typing the essay into an AI detector feels like the obvious next step.

Don't. Or at least, don't stop there. Here's what the detector score actually is, why it fails exactly the students you most need to get right, and what to look at instead.

Why AI detector scores aren't the answer

An AI detector doesn't know anything about your student. It reads the text in isolation and returns a probability — a percentage chance that a pattern-matching model thinks looks machine-generated. That's a guess about style, not a fact about origin, and the guess is built on patterns that don't hold evenly across writers.

The bias is well documented at this point: formulaic, tightly-structured prose — which is exactly what you get from a strong ESL writer, or a neurodivergent student who was taught to write in five clean paragraphs — reads as "more predictable" to these models, and predictable reads as "more likely AI." Meanwhile genuinely AI-assisted writing that's been lightly edited can slide right through. You end up with false accusations aimed at your most careful, rule-following writers, and a pass for the writing that was actually outsourced.

That's not a tuning problem some future update fixes. It's structural: the tool is scoring how the sentences look, not where they came from. Two documents can read identically and have completely different histories.

What actually holds up: look at how it was written

If the text alone can't tell you, look somewhere the text can't fake: the process. A Google Doc has one already sitting in its version history, whether you've ever opened it or not — every keystroke, every paste, every long pause is in there. That process is much harder to manufacture convincingly than a paragraph of prose is.

Here's the rough order I actually use, starting free and going only as far as a given essay calls for:

• Start with the revision history. Open the doc's version history (or use a free tool like Inspect, which replays it for you) and watch how the essay actually got built. A single paste of 900 words at 11:47pm is a completely different story than three weeks of small edits. Real writing is messy — false starts, deleted paragraphs, sentences rewritten twice. A document that appears whole in one clean burst is the single strongest signal you'll find, and it costs nothing to check.
• Compare it to their own writing. Set the essay next to something you know they wrote — an in-class response, an earlier draft, a timed write. Does the sentence length vary the way theirs usually does? Same tics, same go-to transition words, same relationship between vocabulary and error rate? You're not asking "does this look AI" — you're asking "does this sound like them," which is a narrower, much more answerable question.
• Read it against itself. A long essay that's genuinely one person's work usually holds together — register, argument depth, and voice tend to drift together, not independently. Essays stitched from multiple sources (a model, a paraphrasing pass, a friend's paragraph) often show seams: a register shift in paragraph four, an argument that gets suddenly more sophisticated than the surrounding sentences support.
• Check whether the sources actually say that. If the essay cites research, spot-check a claim or two against the actual source. Fabricated or misrepresented citations — a stat that doesn't appear in the paper, a quote that isn't in the book — are a much cleaner signal than prose style, and they catch a different kind of misuse than either revision history or style comparison will.
• When it still matters, ask them. Nothing replaces a two-minute conversation about a specific choice they made: "Walk me through why you started with this example instead of the one from class." A student who wrote it can usually answer without much trouble. A student who didn't tends to stall on specifics, even when the essay itself reads fine.

None of these five, on its own, is proof. A quiet revision history can mean a genuine one-pass writer just as easily as a paste. A style mismatch can mean a rough day, not a ghostwriter. That's exactly why it's worth having more than one signal before you act on any of them — and why the right output of this process is never a single number.

What this actually looks like

Say a junior turns in an American Literature essay that reads noticeably more polished than her last three. Instead of running it through a detector, you open the doc's revision history: it was built over four sessions across five days, with visible deleted paragraphs and a couple of large paste events — both from a shared class document, easily explained. That's not nothing, but it's not damning either, and now you know exactly what to ask about instead of guessing.

Compare that to a doc that appeared as one 1,100-word paste at 10:52pm the night before it was due, on a topic the student never discussed in class, with vocabulary two grade levels above her usual register. Still not proof by itself — but now you have something concrete and specific to raise with her, instead of a hunch you can't defend if a parent pushes back.

The goal was never to catch someone

It's worth saying plainly: none of this is about building an airtight case to nail a kid. It's about being able to walk into a conversation with something more than a feeling — and being able to walk back out of it having actually been fair, in either direction. A revision history that shows honest, incremental work is real evidence too, and it should count just as much as a suspicious one. Students deserve the chance to be believed when they've done the work, not just suspected when they haven't.

The question isn't "does this look AI." It's "can I document what actually happened here, and am I being fair about it either way."

That's the whole idea behind PaperTrail Academic, the set of Google Docs tools I built after running into this exact problem as a classroom teacher: Inspect replays the revision history for free, no account needed. StyleMatch compares an essay against a student's own past writing, entirely on your device. Verify reads an essay against itself for internal consistency. Verify Citations checks whether cited sources actually say what the essay claims. Oral Defense lets a student explain their own choices out loud. Every one of them hands you facts and a question to ask — never a score, never a verdict. The call is always yours.

PaperTrail Academic isn't an AI detector and doesn't produce one. If you need a plain answer for a parent meeting: an AI-probability score is not admissible evidence of anything on its own, and shouldn't be treated as one by your school either.