Safety & support

Companions are fictional AI, not human performers, therapists, or emergency responders. We disclose that before and during a call. Each call starts with an AI notice; continued sessions have a reminder at least every three hours. Current calls are limited to 60 minutes.

The voice agent checks recognized final speech and the latest user text for explicit first-person statements about suicidal thoughts, intended self-harm, or a possible overdose. When final speech triggers the check, the agent interrupts normal turn-taking, publishes crisis links to the call interface, and uses a fixed spoken message directing the user to human support. The language-model path also checks user text before generating a reply. Model instructions prohibit encouragement or instructions for suicide and self-harm, and direct uncertain cases toward human help rather than roleplay.

These checks are not a diagnosis, a clinical assessment, or a guarantee of detection. Speech recognition can be wrong, and indirect language or languages other than English may be missed. Do not rely on a companion to assess your safety. Calls are not monitored by a human emergency team.

We keep aggregate daily counts of crisis referral notifications without conversation text or user identifiers in the counter tables. We test explicit-risk statements and benign/negated phrases, and can refine the protocol as evidence improves. See Privacy for ordinary conversation processing, memory, provider retention, and operational logs.

Hug.Me is for adults 18 and older, not children. Send content or service concerns through Contact. Support email is not an emergency service.