AI Disclosure
Last updated October 17, 2026.
Where AI appears
Pocket Stoa uses a large language model, Anthropic’s Claude, in six places. In each one the model is answering text you wrote, at the moment you send it:
- Your first dialogue with Socrates, while you set up the app.
- The Talk step in each unit of a Path.
- Ask, where Socrates questions a belief you type, or a dilemma or Trolley card you answered.
- Socratic Dialogue, where you question one of twelve fictional characters.
- Premeditatio, which draws three cards for a worry you describe.
- Envy Autopsy, which asks you questions about a post that stung.
Before the first message goes to the model, the app says so and asks for your OK. The rest of the app is pre-written content, not model output. Most of it is written by hand: the quiz, the dilemmas, the Paths, the maxims, the thought experiments, The Last Day and the philosophers’ commentary. Part of the Trolley deck was first drafted with a model, then reviewed and edited by hand before it shipped.
Which model
Anthropic’s Claude (the Claude Haiku models). Your phone never talks to Anthropic directly: it sends your words to our server, which sends them on to Anthropic over HTTPS and returns the reply.
What is sent
What you type, and the conversation so far. With it goes what the model needs to play its part: the brief for the character or for Socrates, which we write; your quiz school (the first dialogue); the unit (a Talk step); the dilemma or card and the side you chose (Ask); and, in Socratic Dialogue, the lines that won your earlier rounds with that character. Your name, email and account id are never sent.
What is kept
Your dialogues (the first one, Talk steps, Ask and Socratic Dialogue) and your Envy Autopsies are stored with your account on our server, so you can come back to them. A Talk step’s first answer and your Premeditatio cards are also kept on your phone. Deleting your account deletes all of it, and Export my data includes it.
Anthropic does not train its models on what the app sends, under its commercial terms, and keeps it only for a limited time, for safety checks. We don’t train any model on your conversations, and we don’t send your text to analytics.
The characters
In Socratic Dialogue the model plays one of twelve fictional characters, each defending a belief. Their personality, biography, belief, and the contradiction that would make them concede are written by us and kept on our server. The model is an improv actor with a tight brief. A second, small call checks whether the character has really conceded the central belief, not just a side point; that is what ends the round and moves your radar.
Reporting a reply
If a reply is offensive, wrong in a harmful way, or just off, press and hold it and choose Report. It disappears from your screen, and the reply comes to us with your account, so we can fix the brief or the filter behind it.
The floor
Language models make mistakes. They will sometimes misattribute a quote, invent a citation, or give a confident answer that’s wrong. Treat Socrates and the characters as imperfect partners in thinking: sharp enough to be useful, never an authority.