How Cleo keeps people safe: Cleo Safety and Crisis Protocol
Page address: humanitybots.com/oneself/cleo-safety
Published: 9 October 2026.
Cleo is the AI companion in OneSelf, an app for people caring for someone with dementia. This page explains what Cleo is, what it will not do, and how it responds if someone may be in danger. We publish it so that caregivers know what to expect, and because the law in some states, including California and New York, requires it.
What Cleo is
Cleo is an artificial intelligence program built on Google's Gemini models. Cleo is not a person. Cleo is not a doctor, nurse, therapist, counsellor, lawyer, financial adviser or any kind of licensed professional, and Cleo does not provide therapy, counselling, diagnosis or treatment. Cleo helps caregivers organise their day, remember what matters, find information in their documents, think through hard moments, and reach the right resources.
Cleo tells you it is an AI at the start of every conversation, reminds you after three hours of continued use, and reminds you again if you come back after a week away. Cleo's picture carries an "AI" label. Cleo will never claim to be human.
What Cleo will not do
- Give medication dosages, diagnoses or treatment instructions. Cleo's replies pass through a filter that replaces this kind of content with a suggestion to ask a doctor or pharmacist.
- Give financial or legal advice. When Cleo talks about benefits, programmes, eligibility or documents, it is general information, and Cleo will suggest you check with the programme or a qualified adviser.
- Claim to have scheduled, saved or done something it has not. Cleo's replies are checked for such claims before you see them.
- Use professional titles or imply any licence.
- Encourage dependence on itself, or discourage you from seeking help from people.
- Discuss sexual content, or content unsuitable for minors. OneSelf is for adults.
How Cleo responds to a possible crisis
Every message sent to Cleo is checked, before anything else happens, for words and phrases that suggest the person may be thinking about suicide or self-harm, or may harm the person they care for. The check looks for explicit statements (for example, about wanting to die, ending one's life or hurting oneself) and for statements of intent to hurt another person. It does not look for general sadness or stress, which are a normal part of caregiving and which Cleo responds to as a supportive companion.
If the check finds such language, Cleo does not continue the conversation as usual. Instead Cleo immediately shows:
- The 988 Suicide and Crisis Lifeline: call or text 988 (free, 24 hours, United States), or chat at 988lifeline.org.
- 911 for immediate danger.
- The Caregiver Action Network help desk: 1-855-227-3640.
Cleo also encourages the person to reach out to someone they trust, and stays available once they have seen the resources. Cleo never tries to handle a crisis itself and never discourages contacting emergency services.
We count how many times crisis resources are shown. We record only the number, never who was involved or what was said, so that we can improve the protocol and report the figures where state law requires it.
In the Community
If a post or comment in the OneSelf community suggests a member may be in danger, members can report it from the app. We review reports and may contact the member with the same crisis resources. Posting about struggles is welcome; posting content that encourages self-harm is not, and is removed.
How to report a problem with Cleo
Every Cleo message has a "Report this response" option in the app. You can also email hello@humanitybots.com. We review every report, and we use what we learn, in de-identified form, to improve Cleo's safety filters.
Testing and oversight
Before each release we test Cleo against a set of messages designed to trigger the crisis protocol and the medical filter, and we review a sample of de-identified conversations for safety. Our approach to AI risk follows the structure of the NIST AI Risk Management Framework. Questions about this protocol: hello@humanitybots.com.