​

What we won’t do

Commitments we make about how we treat users, content, and data.

These are commitments. They define what we will not do, regardless of what becomes technically possible later and regardless of the easier path. If we ever break one, we owe users a public explanation.

​

We won’t host content depicting minors

Not under any creative framing. Not under any “clearly fictional” exception. Not at any age-statement workaround. See Prohibited content.

​

We won’t host characters of real people

No celebrities. No private individuals. No fictional characters that are clearly modelled on a specific real person. See Prohibited content and Intellectual property.

​

We won’t sell your chats

Your private conversations with characters are not for sale. They are yours.

​

We won’t sell or volunteer user data

User data isn’t for sale. We don’t share it with third parties beyond the operational vendors listed in our Privacy Policy. Where law requires us to respond to legal process, we follow the process documented in Contact.

​

We won’t tell other users who chatted with a character

Aggregate stats may be visible. Usernames, chat content, or any identifying information about the people who interacted with a character are not.

​

We won’t black-box moderation

Moderation runs as a hybrid system: automated classifiers handle the volume, human reviewers handle the cases the classifiers flag or can’t resolve. The split is described in detail on How moderation works. When you appeal a decision, a human reviews it — not an algorithm, and not the same person who made the original call.

​

We won’t hide policy changes

This site stays current as our policy evolves. The “Last modified” date at the bottom of each policy page is authoritative; readers checking back see what’s actually in force.

​

We won’t use chat content to train models without consent

Your conversations are not used to fine-tune or train AI models unless you explicitly opt in. Operational processing (running the conversation through classifiers and the generation model) is not training.