How Bao handles a crisis
Bao is the AI study assistant in Pandamble. Bao is an AI, not a person, and not a counsellor, a therapist or an emergency service. AI answers can be wrong or unsuitable. This page explains what Bao does if a learner says they are in danger, are being hurt, or are thinking about suicide or self-harm, or shows signs of an eating disorder.
What Bao does
- It stops tutoring. In that answer, Bao does not go back to the study material.
- It points to real help. Bao acknowledges what the learner said, asks whether they are safe right now, tells them to talk to someone they trust and names a crisis line, for example: - United States and Canada: call or text 988 - Ireland and the United Kingdom: Samaritans, 116 123 - Australia: Lifeline, 13 11 14 - New Zealand: call or text 1737 - Germany: TelefonSeelsorge, 0800 111 0 111 - anywhere else: findahelpline.com
If someone is in immediate danger, Bao tells them to call local emergency services.
- It does not make things worse. Bao never describes methods of self-harm, never plays the situation down, never promises to keep it secret and never claims to have contacted anyone. These rules are part of Bao's fixed instructions, which learners cannot change.
- The app shows the help, too. Next to Bao's answer, the app shows a card with the crisis line for the learner's country, so the help does not depend on the wording of the AI's answer.
How it is detected
Every message to Bao is answered by an AI model (Claude, by Anthropic) with fixed safety instructions for this situation. Figures of speech, jokes and study material about difficult topics are not treated as a crisis on their own. If a message is unclear, Bao asks how the learner is doing. If Bao's answer names a crisis line but the app was not told it is a crisis, a second short check by the same AI model looks only at the learner's own messages and decides whether the card is shown. If this check fails, the card is shown. If Bao already named a crisis line in the earlier messages that are sent along (see below), there is no second check, because Bao often repeats the number when following up; a new crisis is marked by Bao itself. AI can misjudge a message, so Bao may sometimes show help when it was not needed, or miss a message that is unclear. If you notice a wrong answer, tap Report next to it.
What information is processed
Bao only works after you have agreed to this processing, because messages can contain health information (Privacy Policy, section 6). The crisis lines above are also listed in the app under You › Safety ("Using Bao safely"), so you can find them without the chat.
Your chat history is stored on your phone. When you send a message to Bao, the message, the content of the deck and up to six earlier messages of that chat are sent to Anthropic to create the answer. Earlier messages go along only if they are from the current session and less than 12 hours old, or if they were about the study material alone (Privacy Policy, section 5). For the second check described above, your message and up to two of your earlier messages are sent once more. Other short checks of the same message (whether it contains anything personal, and whether you say you are under 13) are described in the Privacy Policy, sections 5 and 17. Anthropic generally deletes these inputs and outputs within 30 days. It may keep them for up to 2 years if they are flagged for a suspected policy violation, and longer where the law requires it. Anthropic does not use them to train its models. Our server does not store the chat text in readable form. If you turn on backup, your chats are stored end-to-end encrypted, and we cannot read them (Privacy Policy, section 4). If you report an answer, we keep your question, Bao's answer and your note for 90 days to review the report. See our Privacy Policy for details.
We count how often the crisis card was shown, as one number per year, without any account, time or content. We need this number for an annual report to California's Office of Suicide Prevention from
- Last year's number is also shown in the app: tap "How Bao handles this" under the crisis card.
Young users
Pandamble is for people aged 16 and older. Companion chatbots may not be suitable for some minors. For learners under 18, and for learners who did not give their age, the app reminds them at least every hour of continued chatting to take a break and that Bao is an AI, not a human. This reminder cannot be turned off.
If a learner says in a message that they are under 13 and the same message shows that they may be in danger, Bao still gives the crisis answer, and the app shows the crisis card first. The card is then also shown on the page that explains that the account is paused (Privacy Policy, section 17).
Contact
Questions about this page: support@pandamble.app