Clues, not answers: how an AI study tutor should behave

A student stuck on question 4 at nine o'clock at night wants one thing: the answer. Any tool that supplies it has solved tonight's problem and created tomorrow's, because the student has not learned how to do it.

Schools are right to be wary of tools that do the work for a student. The useful question is what a study companion should do instead.

A ladder of help

A good tutor does not jump to the answer. It starts by asking where the student is stuck, then offers a small hint, then a bigger one, then an example of a similar problem. At each step the student does the next piece of thinking.

• Ask what the student has tried and where it stopped working.
• Give a clue that points at the idea, not the result.
• Give feedback on the student's own attempt, naming what is right and what is missing.
• Set a fresh problem of the same kind, to be answered without help.

Mastery needs a check the student does alone

Getting an answer right with three hints is progress, but it is not the same as understanding. In Schanion a topic counts as mastered only when the student passes an independent check with no hints. Practice earns XP, which rewards effort, and the two numbers are kept apart on purpose so that effort can never be mistaken for understanding.

What it will not do

The companion will not hand over the model answer on an assessment task, and a teacher can lock a task so that the companion's help is limited while it is open. It refuses requests to ignore its rules, and it tells a student that it is an AI, not a person.

A plain note on what is built

The companion in Schanion today is a scripted guide, not a language model. It can say only what has been written for it, which keeps it predictable and safe, and also limits how well it handles questions outside its content. Adding a language model would be a separate decision that needs its own safety review. You can see the full list of what is built and what is not on the About page.