Is an AI detector score enough to accuse a student of cheating?

On its own, no, and the vendors say so: the percentage is an indicator meant to start a conversation, not a finding of misconduct. The score is a statistical guess about how predictable the finished words are, produced by a system that never saw the essay being written. Most schools that have written an AI policy require corroborating evidence — a process record, a conversation, an in-class writing sample — before anything goes on a student's file.

Why do AI detectors flag essays that students actually wrote?

Detectors score how predictable a text is. Writing that is plain, formulaic, or heavily edited toward a rubric scores as predictable, because that is what the model expects. This hits multilingual and English-learner writers hardest, since a smaller and more careful vocabulary looks statistically flat. It also hits students who were taught a rigid structure, students who use grammar assistants, and students who write short declarative sentences. None of that is cheating; it is writing that happens to be easy to predict.

What evidence actually settles whether a student wrote an essay?

Evidence about how the text came into existence, rather than about how the finished words read. A drafting timeline showing sentences appearing, being deleted and rewritten over time. The size of each insertion, since typing arrives a few characters at a time and a pasted block arrives all at once. Blocked paste attempts with timestamps. Window focus losses. A student who wrote the essay can also usually explain a specific choice in it. These are observations with times attached, which is why they survive being argued with.

An AI detector flagged my student's essay

You have a number on a screen and a student who says they wrote it. The number cannot tell you who is right, and neither can rereading the essay for the fourth time. This is what the number actually is, what to do in the next twenty-four hours, and what evidence ends the argument in either direction.

What the score actually measures

An AI detector reads finished text and estimates how predictable those words were. If the next word is usually the obvious one, the score goes up. That is the whole mechanism. It is a statement about the statistical shape of the sentences, not about the history of the document, and the detector never observed a single second of the writing.

This matters because "predictable prose" and "machine-written prose" are overlapping categories, not the same category. Plenty of human writing is predictable on purpose. A student following a five-paragraph template produces predictable prose. A multilingual student writing carefully in a second language, choosing safe vocabulary and simple constructions, produces predictable prose. A student who ran a grammar assistant over their draft produces prose that has been smoothed toward the average. All of those land in the same statistical neighbourhood as generated text, and the detector cannot separate them, because the separating information is not in the words.

The vendors know this. Turnitin's own documentation describes its AI indicator as something to prompt a conversation rather than a determination of misconduct, and several large universities turned their detector off rather than keep managing the appeals. That is not a claim that no student ever pastes from a chatbot — plenty do. It is a claim that the score cannot tell you which student is which, and a false accusation is expensive: it lands on a real teenager who now believes the school will not believe them.

What a percentage does not tell you

It is not a probability that the student cheated. "98% AI" does not mean a 98% chance of misconduct. Depending on the tool it can mean the share of sentences scored as machine-like, or a confidence band, and the label differs between vendors and between versions of the same vendor.
It has no time in it. Nothing in the score says when the document came into existence, in what order, or across how many sittings.
You cannot reproduce it. Run the same essay again next term, after the vendor updates its model, and the number can change. Evidence that changes on its own is hard to defend in a meeting.
It is unevenly wrong. The errors are not spread randomly across your class. They cluster on the students who can least afford them, which turns a technical limitation into an equity problem.

What to do in the next day

Do not open with the number. The moment a student hears a percentage, the conversation becomes about the software. Open with the essay: ask them to walk you through how they got to their argument, why they cut the paragraph that clearly used to be in there, where they found the second example.
Ask for what they already have. Version history, notes, an outline, a photo of a scribbled plan, a draft they mailed to themselves. None of it is conclusive alone, and it is worth asking for what exists before anything new gets produced.
Write down what you observed, not what you concluded. "The essay uses two sources that were not in our packet, and the student could not say where they came from" is a note that survives a parent meeting. "The detector said 87%" is not.
Offer a clean second chance to demonstrate. A short supervised piece on a related question, recorded, where the drafting itself is the evidence. This is the step that resolves the case, because it produces the one kind of proof the original essay is missing.
Say what the school's policy is, out loud. If your policy requires corroboration beyond a detector score, tell the student that in the first meeting. It changes the temperature immediately.

Why a process record settles it either way

Everything above is a workaround for one missing thing: nobody watched the essay being written. A process record is that missing thing, captured automatically while the student writes.

When the writing is recorded, the drafting timeline is the evidence. You see the first sentence go in, get deleted, come back in a different order. You see a nine-second pause before the hard clause and a two-minute pause where the student went to look at their notes. You see that nearly every character arrived a few at a time, and that the largest single insertion was eleven characters. That shape is hard to fake and easy to read. A student who was accused wrongly is cleared by it in about ninety seconds, which is the part teachers tend to underestimate: the same record that shows pasting is the record that clears the kid who did the work.

It cuts the other way too. If the record shows a 1,400-character block arriving in one edit at 10:14, or text arriving at a flat, even cadence with almost no deletions, you are no longer arguing about tone and vocabulary. You are describing an event with a time on it, and the conversation is about that event.

Being fair to the student is not a side benefit here, it is the point. A student who is told "the software says you cheated" has nothing to answer with. A student who is told "here is the recording of you writing it" either has the best possible defence or knows exactly what you are looking at. Both of those are better than a standoff.

Where WrittenLive fits

WrittenLive records the writing process and replays it. That is the entire product. You type an assignment title and get two links: one to send your class, one private dashboard for you. Students open the link, type a name, and write — no accounts, no installs, fine on a Chromebook or a phone. Paste, drop and drag-in are blocked and counted. Every insert, delete, cursor jump, focus change and idle gap is timestamped, so afterwards you get a scrubbable keystroke replay, a class table, and a one-page printable authorship report per student.

It makes no AI-detection judgment. There is no model in it scoring anything, no percentage, and no verdict — every figure on the report is something the editor observed while the student typed. That is deliberate. A guess about finished words is what put you in this situation; another guess would not help. For the difference between the two kinds of claim, see how to prove a student wrote it.

The limits, said plainly

A recording of the next essay does nothing for the essay already in dispute. If the flagged piece was written at home in a document with no meaningful history, no tool can reconstruct what happened, and it is better to say so than to reach for a number that sounds like certainty. A student can also retype generated text from a phone, and the record will say "typed", because it was — transcription tends to look unusually even, with few deletions and few long pauses, but that is a pattern worth a question, not a finding. And a flag is a reason to look, never a decision. Take the report to the conversation; do not send it instead of having one.

Made by a developer after teachers asked for it. I am not a teacher. I built this after reading thread after thread of people being handed a detector percentage and no way to check it. Free for one class this school year, no card anywhere in the product.

Give the next essay a record

Or look first: the public demo has one clean session and one with paste attempts, a large insertion, and machine-speed bursts.