How AI Works in Selm

1. What the AI does in Selm

1.1. The AI co-author only performs actions you start: it builds a canon brief, pitches plot ideas, drafts an outline, writes chapters, continues text, suggests next beats, edits a selected passage, checks consistency with the canon and suggests titles.

1.2. Every paid action shows its cost in sparks first. AI edits can be undone, and earlier versions of a chapter are kept in its version history.

2. Which models are used

2.1. Text is generated by Anthropic (Claude) or OpenAI language models, depending on the Service’s current settings. For Pro subscribers the model gets more time to reason, or a stronger model is used.

2.2. Story audio is created by an OpenAI speech-synthesis model. It is a synthesised voice, not a recording of a person reading.

2.3. Models may change as the Service develops.

3. What data the AI receives

3.1. To work on your story, the provider of the model currently writing the text (Anthropic or OpenAI) receives:

• the story’s title, premise, branch point, format, language and 18+ flag;
• the outline and summaries, or the openings and endings, of earlier chapters;
• the current chapter, within a size limit;
• your instructions, wishes and the selected passage;
• the canon brief for the work.

3.2. For audio, the text of the published chapter is sent to OpenAI.

3.3. Not sent: your name, username, photo, Telegram ID and payment data. Requests carry only a pseudonymous identifier — a hash — so the model provider can fight abuse.

4. What is stored

4.1. Selm doesn’t keep a separate log of AI requests and answers. Generated text is saved only where you put it: in your chapter and its version history.

4.2. For cost accounting and abuse prevention we store technical data for each request: kind of action, model, size in tokens, cost, time and outcome.

4.3. Requests are sent to OpenAI with storage on OpenAI’s side turned off. Under OpenAI’s API policy, data may be retained for up to 30 days for abuse monitoring and is not used to train models.

4.4. Under Anthropic’s commercial API terms, request data is not used to train models and is kept for a limited time for abuse monitoring, then deleted.

5. The canon brief

5.1. The brief on a work is built from open sources: Wikipedia, TVmaze, Shikimori and catalog data. The AI relies on these sources and should not invent what isn’t there.

5.2. The brief is shared by everyone writing about the same work, and its sources are linked in the app. Briefs are free, but the number of new briefs per day is limited.

6. The AI’s limits

6.1. The AI can make mistakes: mix up canon details, repeat itself, distort facts. Check the text before publishing.

6.2. The AI is set up not to produce explicit sexual content. There is no automated moderation: violations are handled through user reports.

6.3. Don’t use the AI to write about real people doing things they didn’t do.

7. Stories written with AI

7.1. Stories on Selm may be written with AI fully or in part. The author who publishes the text is responsible for it.

7.2. Rights to text generated at your request are covered by the Terms of Use.

8. Limits

8.1. AI actions spend sparks: free ones (20 a day), Pro ones (150 a day) or purchased ones. The Service has a shared daily AI budget: once it’s used up, free actions pause until the next day, while Pro subscribers and holders of purchased sparks keep working within a higher limit.