Keep the learner in the work.
Socrates should create room for a person to notice, explain, test, and revise their own understanding.
Explore the learning loopPrinciples
AI can make learning feel faster. Our job is to make it more honest, more legible, and more likely to last.
Make uncertainty useful.
A hesitant answer is not a failure state. It is often the clearest signal for what question should come next.
Read the researchShow the path, not only the result.
Useful learning systems keep enough context to reveal where an idea began and what it now connects to.
See knowledge mapsUse care with confidence.
When a tutor is uncertain, it should say so plainly and help a learner examine the assumption rather than cover it up.
Talk with usHow we choose a useful direction.
Every educational tool changes the distribution of attention. It can turn a hard question into a shortcut, or it can help someone stay with the question long enough for understanding to form.
We build Socrates toward the second possibility: a tutor that is rigorous about evidence, generous with context, and unwilling to confuse a fluent answer with a learned idea.
A working framework
Clear constraints make better questions possible.
We use these principles to judge product decisions, research directions, and the way a new capability is introduced to a learner.
Read our notesWhat we will not do
Lines we hold even when the model gets more capable.
A set of principles is only useful if it is willing to refuse things. The following commitments tell you the kinds of behaviour Socrates is designed to avoid, even when an intervention would otherwise be persuasive.
Refuse
We will not paper over a gap with fluency.
If a question goes beyond what the tutor can support, it should say so plainly and offer a better next step—never a fluent sentence that pretends to know.
Refuse
We will not become an answer-generating machine.
Features that would let a learner skip the work they are here to do will be deferred or refused, even when they look attractive in a demo.
Refuse
We will not design for compulsive use.
There are no streaks, no infinite feeds, no dopamine loops. The interface gets quieter as the conversation deepens, not louder.
Refuse
We will not use your sessions to train future models.
Your thinking belongs to you. The privacy page describes exactly what is stored and for how long. We export, delete, and forget on request.
How we decide
A short description of our decision process.
When a new capability is proposed, we walk through the same five questions. The answers are written down, attached to the change, and made available in the announcement when the feature ships.
01Does it return the work to the learner?Or does it do the work for them?
02Does it preserve the path into the question?Or does it collapse a sequence into a single turn?
03Is the cost of a mistake legible?Can the learner see where the tutor was wrong?
04What is the worst plausible misuse?And how does the design respond to it?
05What does the learner keep?What is private, what is portable, what is exportable?