Suicide and Self-Harm Prevention Protocol

If you are in immediate danger, feeling suicidal, or experiencing a mental health crisis, contact emergency services or a crisis line now — do not wait for the app to respond.

• United States & Canada: call or text 988 (Suicide & Crisis Lifeline), or call 911
• United Kingdom: call 111, or 999 in an emergency; Samaritans 116 123
• European Union: call 112
• Elsewhere: check your local crisis service

Per Terms of Service section 5.1, our AI Companions are prompted, through their underlying system instructions, to respond to expressions of self-harm or suicidal ideation with care — taking the disclosure seriously and pointing the user toward real help, such as 988 in the US. The Companion stays in character while doing this. This is a conversational safeguard built into how the Companion is prompted to respond, not a dedicated automated detection system, and not a substitute for professional help.

The Companion responds in persona, as described in Section 2. The exchange is captured in our standard application logs. This is not currently a dedicated, structured safety-review pipeline — it is the same general-purpose logging used elsewhere in the product, which may be reviewed if a concern is raised.

Do not wait for or rely on the AI to recognise or handle a crisis. AI systems can make mistakes, fail to detect certain situations, experience technical interruptions, or respond in ways that are not appropriate for an emergency. If you or someone else may be at immediate risk, seek help from a qualified mental health professional, a trusted person, a crisis service, or local emergency services immediately.

We do not currently run a classifier dedicated to detecting suicidal ideation or self-harm language; the current behaviour depends on the underlying language model's own judgement in the moment, guided by the system instructions described in Section 2. It will miss some genuine expressions of distress. It is one safety consideration among several, not a substitute for human care, professional support, or emergency services, and you should not rely on it as your sole safety net. See Terms of Service sections 3.2 and 4 for the full set of warnings about what this product is and is not.

Where the Companion points a user toward help, the resources referenced are limited to publicly available crisis lines and, where relevant, general information about bereavement support. We do not refer users to a specific clinician, do not share conversation content with a third-party provider as part of a referral, and do not condition continued access to the Service on accepting a referral.

We are aware that a more robust approach — dedicated crisis-language detection, an out-of-character interruption, a persistent on-screen crisis resource panel, and a structured safety-review process — would better match user expectations and regulatory guidance in this area, including obligations that may apply under laws such as California SB 243. This Protocol will be updated if and when that capability is built.

We will not weaken the protections described in Sections 1 through 3 without at least 30 days' advance notice, consistent with Terms of Service section 20.