A student hands in work that was partly written by a chatbot. You notice. Now what.
The instinct is to escalate, or to ignore. Both cost you the learning. There is a calmer third path, and it takes about five minutes.
The three-step protocol
The protocol has three steps: declaration, gate check, and re-prediction. It sits inside an active-inference frame of teaching (Class E, see Parr, Pezzulo, Friston 2022 for the underlying generative-model account), and it is designed so the student stays the agent of their own learning rather than the object of a discipline routine.
Step 1: Declaration
Ask, once, plainly: "Tell me which parts you wrote, which parts a tool wrote, and which parts you edited after the tool wrote them."
Do not ask "did you cheat." That question invites a defensive answer and teaches the student nothing about the boundary you actually care about. The declaration question teaches the boundary directly. It also matches the AI-authorship fence the class already uses on every artifact (Class C, this is how the fence is wired into our classroom template).
Accept the answer. If it changes under gentle re-asking, accept the revised version. The goal is an accurate map, not a confession.
Step 2: Gate check
Now go to the gate the assignment was meant to open. Every assignment in this method has a gate: a specific thing the student should be able to do afterward that they could not do before. The gate is not the artifact. The artifact is evidence about the gate.
Pick one item from the gate and ask the student to do it, out loud, with you, using a fresh example. If the assignment was "explain why this circuit lights up," pick a different circuit and ask them to walk through it. If the assignment was "argue a position from two sources," hand them a third source and ask what it changes.
You are not testing whether they memorized. You are checking whether the prediction the assignment was supposed to install actually got installed. This is the check step in the predict-then-check loop the class already runs on ordinary lessons.
Step 3: Re-prediction
If the gate holds, the learning happened. Note the tool use on the artifact per the fence, and move on. The student used a tool, declared it, and can still do the thing. That is a workable outcome.
If the gate does not hold, you have information. The chatbot produced an artifact that outran the student's own model. That is a prediction error, in the technical sense, and it is exactly the moment the student can update. Give them a short re-prediction task: same gate, new example, done in the room, no tools. Grade the re-prediction, not the original artifact.
This keeps the incentive pointed the right way. The student learns that the artifact is not the point, the gate is the point, and using a tool is fine as long as the gate still opens under their own hand.
What this protocol is not
It is not a lie detector. It is not a way to catch students. It is not a claim that chatbot use is harmless or harmful in general. It is a way to keep the learning intact when a tool has been in the loop, which, in 2026, is most of the time.
It also is not therapy or a discipline meeting. If something else is going on with the student, that is a separate conversation, with the people qualified to have it.
Why this works inside our frame
Under the active-inference view of learning we teach here, a student's understanding is a generative model they can run on new examples (Class E). A chatbot can produce an artifact that looks like a good model output without the student's own model having updated at all. The gate check is a direct probe of the student's model. The re-prediction is a supervised update. Neither step requires you to know what the chatbot did, or to prove anything about it. What you need to know is whether the student's own model can carry the load the assignment was designed to build.
That is why the protocol is calm. You are not adjudicating the tool. You are checking the gate.