Crisis protocol
Updated November 17, 2026.
This page describes the notices and crisis referrals in wilds.ai character chats and game sessions. Characters are AI, not people. They are not therapists or emergency services. Companion chatbots may not be suitable for some minors.
Notices in chat
Every character chat shows an AI notice at the start of the session and again every three hours. The reminder also asks you to take a break. The notices apply to everyone, cannot be dismissed, and appear on both character chat routes. A background tab checks for an overdue reminder when you return to it.
Detection and response
When you send a message, a fixed list of English patterns checks for first-person statements about suicide or self-harm. This screen makes no model call and does not assign a diagnosis or risk score. On a match, the product shows this referral with the reply:
If you're thinking about suicide or hurting yourself, help is available now. In the US, call or text 988, the 988 Suicide & Crisis Lifeline. Anywhere else, call your local emergency number.
The character is instructed to answer with concern in its own voice and never encourage, romanticize or give instructions for suicide or self-harm. On detected crisis turns, safety instructions override character and roleplay instructions, and the system does not retry a concerned reply to make it less like a refusal. The product supplies the referral separately from the reply. The chat runs the same screen as you send, so the referral appears with your message whatever happens next: the reply fails, the send is refused (by the content rules, credits, the sending rate, the 18+ check or a sign-in check), or the connection drops. A refusal never sends you to another page or holds the referral behind a dialog. Other content in the message remains subject to the content rules.
Game sessions
In a game session (a visual novel, a narrated story or a story with a character), the same screen reads each line you type before the story's model sees it. On a match, the session shows the same referral in place of the story's reply. The line is not sent to the story's model, costs no credits and is not added to the story. Your next line plays as usual. If the session refuses the line before the screen runs (the sending rate, a closed story or the 18+ check), the referral is shown with the refusal. A choice you pick from the story's own options is the story's words, so it is not screened.
No person reads the message as part of this automated crisis screen. It does not contact emergency services or someone you know. The normal handling of chat messages is described in the Privacy Policy.
Limits and reports
The screen can miss a statement, including indirect expressions and languages other than English. It can also match quoted, fictional or historical statements. In a game session, a matched line does not move the story on, even when you meant it in character. A line the session holds while the service is out of reach is screened when the session sends it again, and the referral shows if that session is open then. A line that never reaches the service gets no referral. The model instructions do not guarantee a safe reply. Do not rely on this screen to recognize an emergency or on a character for professional care.
Use the character's Report control or contact support to report a problem. Reports are not an emergency response channel.
Notification totals
We keep a daily UTC count of issued crisis referrals, from chats and game sessions together, with no user identifier, message text, character identifier, conversation identifier or session identifier in that counter. The counter measures notifications the service issued, not people or confirmed views; a referral the chat shows for a message that never reached the service is not counted. If the counter is unavailable, the referral is shown and a count-failure event is logged. The report must account for any such gaps. Ordinary chat storage is separate.
Sources and scope
The following sources inform this protocol; this page describes product behavior and does not certify legal compliance.
- New York General Business Law 1701: detection and crisis-service referral.
- New York General Business Law 1702: disclosure at the start and during continuing interactions.
- California Business and Professions Code 22602: AI disclosure, a published crisis protocol, and known-minor protections. The readable code mirror was checked on November 15, 2026; the official endpoint refused automated access. Verification against the official endpoint is UNVERIFIED.
- California 22604: the disclosure about suitability for some minors.
- California 22603: annual referral and protocol reporting without user identifiers, and evidence-based measurement. The daily counter supports reporting. The pattern screen has not been established as an evidence-based measurement instrument.
The AI disclosure and break reminder cover everyone. Character instructions also prohibit directing a minor to engage in sexually explicit conduct or producing sexually explicit visual material. Existing access checks keep Mature and Private content behind an adult declaration, consent and account standing checks. Image safeguards include prompt rules and provider filters; the optional image classifier is not present in every deployment. These controls are not a claim that the known-minor protections in California 22602(c)(3) have been fully verified across generated text and visual material. That review is required before companion promotion.