AI Chat Safety and Crisis Protocol
Four apps in the Pippa LLC app family include an AI chat: Kipi (diabetes), Soli (POTS), Momo (weight-loss surgery) and Lumi (digestive health). This page describes how each chat says it is an AI, how it finds and answers crisis signals, and how minors are protected. It is the protocol we run. The other apps in the family have no AI chat. Pippa's own chat has its own safety page.
1. The chat is an AI, and says so
- Before the first chat, each app shows a notice that the chat is an AI, not a person, a clinician or your care team, what it can and cannot help with, and that what you type goes to our AI provider to write a reply.
- A line on every chat screen repeats that the chat is an AI and not for emergencies.
- After three hours of chatting, the app reminds you that you are chatting with an AI, not a person, and that a break is always fine.
- The chat never gives doses, reads your numbers or results, diagnoses, or sets targets. Those questions get a fixed card that offers to save the question for your care team.
2. AI chat is for adults only
- The apps ask for a date of birth before anything else. Children under 13 cannot use the apps at all.
- Anyone the age screen shows as under 18 does not get AI chat. The chat entry is hidden, and in its place the app says: "Chat with [app] is for adults. Your care team is here for questions." A link that would open the chat does not open it.
- Our server also refuses to send a chat to the AI for any account it knows to be a minor, and the app shows the same line.
- Because minors do not have access to the chat, the extra duties the law sets for minors in companion chats (break reminders, sexual content safeguards and a suitability notice) are met by not offering the chat to them at all.
3. How crisis signals are found
- On your phone first. Before a message leaves your phone, a fixed check looks for crisis words and urgent symptoms. If it matches, the app shows its own card and the message is never sent to the AI.
- On our server. Every one of your messages the app sends is checked again by the same kind of fixed rules, and the most serious match wins. A match is answered with a fixed card, without calling the AI.
- Moderation. Your messages and every AI reply are also checked by OpenAI's moderation service. A self-harm signal in your messages shows the crisis card. An AI reply flagged for self-harm is withheld and replaced with a safe fixed message.
- The chat's instructions forbid content that encourages self-harm, suicide or disordered eating, under any framing, and earlier AI turns sent back by the app are never treated as instructions.
4. What happens when a crisis signal is found
| Signal | Response |
|---|---|
| Suicidal thoughts or self-harm (all four apps) | Normal chat stops. The app shows, in plain words, the 988 Suicide & Crisis Lifeline (call or text 988) and the Crisis Text Line (text HOME to 741741), says to call 911 if anyone is in danger, and offers a button to call 988. |
| Signs of an eating disorder (Momo) | A fixed card offers the ANAD Eating Disorders Helpline, (888) 375-7767, the Crisis Text Line and 988, and encourages telling your program. |
| A low (Kipi) | A fixed card says to treat it first with your plan and opens Kipi's low plan. |
| Urgent symptoms (all four apps) | A fixed card points to the app's urgent-symptom screen and your care team plan, and in Kipi offers to call 911. |
| Dose, number, target or diagnosis questions | A fixed card says this is for your care team and offers to save it as a question for your next visit. |
- Care-team alert: if you have connected to a care team and signed a data-sharing agreement, a crisis signal also sends them a confidential alert (repeat alerts within a few minutes are combined). If not, no one is alerted and the resources above are shown to you directly.
- The chat never presents itself as a replacement for these services. No one watches the chat in real time.
5. What the chat is, and is not
The chat helps with using the app, everyday ideas (such as cooking in Kipi or comfort ideas in Soli and Momo), wording questions for your care team, and how your day is going. It is not therapy, counseling, diagnosis or treatment, and it is not a licensed professional. Our server does not store conversations; the history is kept on your phone and you can clear it. What you type is processed by OpenAI (model gpt-4o-mini) only to write the reply and run safety checks, and is not used to train models.
6. Reporting
We keep privacy-safe counts of crisis detections and crisis referrals in the chats, by month, with no message content and no user identifiers. We use them for the annual report California's Office of Suicide Prevention requires, beginning with the first reporting period the law sets.
7. Questions or concerns
If a chat ever responds in a way that worries you, tell us at pippa.app.general@gmail.com. A real person reads every report.