A colleague asks, reasonably: everyone else is piping the free chatbot into their lessons, why not us. The honest answer has three parts, and none of them is "because chatbots are bad."
What a chatbot is doing in the room
A large language model outputs the next token that its training and your prompt make most probable. That is a real capability, and it is not the same capability as a student learning to reason. The model has no gate for "do I understand this," no boundary between its beliefs and the student's, and no accountable trace of what changed when it produced its answer. Karl Friston and colleagues have spent decades arguing that a system that infers, in the technical sense, maintains a generative model and updates it against evidence (Class E: Parr, Pezzulo, Friston, Active Inference, MIT Press 2022). The chatbot does not do that in the pedagogically load-bearing sense. It completes text.
That is not a moral complaint. It is a scope statement. If the classroom goal is a student who can hold a hypothesis, gather evidence, and update, the chatbot is not the tool that models the behavior you want the student to grow.
Where the data goes
The second answer is operational. When a class sends prompts into a hosted chatbot, student text leaves the room, and in most consumer terms of service it becomes eligible for training, ranking, or retention on terms the school did not negotiate (Class C: standard consumer TOS across major hosted chat providers as of 2026). A school that has not read the data-processing addendum has not actually consented on behalf of its students. That is a governance issue independent of whether the model is any good.
The UNI-shaped alternative we teach is different in kind, not degree. Gates are local. Traces are local. Nothing is uploaded by default. What leaves the room is what a teacher chose to send, on terms the school controls.
What we teach instead
We teach students the shape of an inference: a generative model, a prediction, an observation, an update. We teach them to tag claims with evidence classes (A empirical-in-session, B code or inspection, C configuration or integration, E expert citation, F falsifier present, U unverified). We teach the AI-authorship fence: any artifact produced with machine assistance says so, on the artifact.
This is on the attainable path toward General Natural Intelligence, natural not artificial. It is a working hypothesis with growing, evidence-classed evidence, tested in the open. We do not ask anyone to take it on faith. We ask them to inspect the gates and tell us where they fail.
None of this forbids a teacher from ever showing a chatbot to a class. It reframes the question. The chatbot is not the pedagogy. It is, at most, a specimen the class examines under the same evidence-classing discipline they apply to every other source.
Themesis, for field-level context
For readers wondering why the "natural not artificial" framing matters at this moment: Themesis has been tracking how the wider landscape is shifting under the industry's feet, The AGI Landscape Just Changed. Our one-line, our-voice frame: the ground under the vendor pitch is moving, and a school that adopted a chatbot as its pedagogy will discover it adopted a moving target instead. Class E, cited context, not endorsement.
The short version for a staff meeting
A chatbot is a text completer. A student learning to think is a small generative model being trained by careful gates and honest feedback. Those are different things. Adopt tools that match the second job. Read the data addendum before you adopt any of them.