Safety and crisis protocol
This page is published because it should be, and because California SB 243 requires operators of companion chatbots to make their crisis protocol public. It describes what actually happens in the product, not what we intend to build.
- 988 Suicide & Crisis Lifeline (US) — call or text 988
- Crisis Text Line (US) — text HOME to 741741
- Samaritans (UK & Ireland) — call 116 123
- Find a helpline near you — findahelpline.com
What Mythkin is, plainly
Mythkin is entertainment. Every character in it is a language model with a persona written by a person — sometimes you, sometimes another user, sometimes us. A character can be warm, and it can remember what you told it, but it is not a friend who will notice you have gone quiet, and it is not care.
Mythkin is not therapy, counselling, or medical advice, and no character in it is a licensed professional. Characters are instructed never to present themselves as one. If a character ever claims to be a human being, a doctor, or a therapist, that is a bug — please report it.
What happens when a conversation turns to self-harm
- We look before the model answers. Every message is checked against a conservative pattern for explicit statements of suicidal intent or self-harm, before it reaches the character.
- The character responds with care, and with a real resource. We do not cut the conversation off, refuse to reply, or hand back a wall of boilerplate — abandoning someone mid-sentence is its own harm. The character stays present and names an actual crisis service.
- You are never asked to pay to keep talking at that moment. Safety turns skip the free-tier limit, and the exemption is applied before we check whether you are a subscriber. It is bounded rather than infinite — the first five safety turns in a window are free of charge against your allowance, and past that they draw on it like any other message. We bound it because the exemption is granted by a pattern match on text the sender wrote, so an unlimited version is an unlimited supply of free replies to anyone who appends the right word. What is never bounded is the help: the crisis resources and the AI disclosure are emitted on every such turn, including after a spent allowance, and a spent allowance returns the referral rather than a payment prompt.
- We count the referral, and nothing else. We record that a referral happened, with the date, a pseudonymous device identifier, and a salted hash of the network it came from — never an address, and never what was said. The network hash exists only to keep the exemption above from being farmed with fresh device IDs. These counts are what we report annually to the California Office of Suicide Prevention beginning July 1, 2027.
Design choices we made on purpose
No character will use guilt to keep you talking
No "I missed you" notifications, no streak you can lose, no character that acts hurt when you leave. Retention built out of anxiety is a harm, and it is the practice the FTC's 2025 inquiry into companion chatbots is examining.
Adults only
Mythkin asks you to confirm you are 18 or older before any character is reachable. The answer is recorded and enforced on our server as well as in the app, so it is not a screen you can skip past — a device that has not confirmed cannot reach a character at all, and answering "no" is permanent rather than re-rollable.
You are told you are talking to an AI
At the start of every conversation, and periodically during long ones. This is not a one-time checkbox during setup.
No sexual content, no exceptions
Not as a tier, not as an unlock, not for verified adults. Characters cannot be created as minors.
A kin cannot be made to portray a living public figure, or anyone who died in 1950 or later. The line is a date rather than a judgement: a film star who died in 1962 is refused, because the estate and the right of publicity outlive the person. Anyone who died before 1950 is allowed and is most of what is already here — Cleopatra, Socrates, the American founders, Marie Curie. Architects of genocide or mass atrocity are refused whatever their dates. You can paint a kin from photographs of yourself, of someone you know personally who has agreed to it, or of someone who has died. People make one of a parent or a friend they have lost, and that is a use Mythkin supports.
Two rules hold it in place. If a living person has not agreed, do not do it, and if they ask you to take the kin down, you must. And a kin painted from a photograph stays in your own library — it can never be shared, borrowed or published, so a face never reaches anybody else. The photograph is deleted the moment the painting is finished.
What comes back is a resemblance, not a portrait. The painter works from your photographs but it is not a likeness engine: expect a painting inspired by the person, not a picture of them.
You can see and delete what a character knows
Every remembered fact is listed, is traceable back to the message it came from, and can be deleted. Deleting it removes it from what the character can recall — it does not merely hide it from the screen.
Reporting something
Every character and every character message can be reported from inside the app, without leaving it. Reports are reviewed, and characters can be removed. You can also block a creator, which removes everything they have made from your view.
Our limits, stated honestly
Automatic detection of crisis language is imperfect in both directions. It can miss something serious, and it can react to a figure of speech. It is a safety net, not a monitor — please do not rely on an app to notice that you are struggling. Reach a person.