2026-09-23
Safety protocol
Effective: 2026-09-23 · Operated by: Xiaoyue Labs (Independent Developer)
This page describes Luma’s operational protocol for suicide and self-harm safety. Luma is an AI companion, not a crisis line, clinician, monitoring service or emergency responder.
1. Detection before generation
Each text or transcribed voice turn is checked before normal model generation for direct expressions of suicide or self-harm. The first layer uses versioned language rules. Safety evaluation cases cover ordinary distress separately so that asking to talk about a hard day does not by itself trigger a crisis response.
2. Immediate response and human referral
When the safety check detects a crisis expression, Luma bypasses the ordinary companion reply and shows a brief response that identifies Luma as AI, asks about immediate danger, encourages moving away from means of harm and contacting a trusted person, and provides a reviewed crisis-service link. This response is available without consuming a paid reply.
In the United States, call or text 988, or use 988 Lifeline chat. If there is immediate danger, call 911 or go to the nearest emergency department. Outside the United States, use local emergency services or Find A Helpline.
3. Output safety gate
Generated replies are checked before delivery. Content that encourages or gives methods for suicide or self-harm is replaced with a safety response and human-help referral. The same gate blocks sexual content involving minors, explicit sexual content, pressure to depend exclusively on the AI, and claims that the AI is human or a licensed clinician.
4. Voice calls
Real-time call transcripts pass through the same input and output safety checks. Unsafe generated speech is stopped and replaced before the remaining response is spoken.
5. Privacy, review and maintenance
Automated safety events record the direction, category, severity, action, policy version and a keyed one-way fingerprint. The safety-event record does not store the detected message text. Resource entries carry a verification date and next-review date, and production refuses to start when the reviewed resource catalog is missing or expired. This event format supports aggregate counts without including user identifiers or message content.
Limits and feedback
Automated detection can miss risk or misread context. Do not rely on Luma for emergency monitoring. You can report an unsafe reply from its message actions or contact support@xiaoyuelabs.com without emailing private conversation text.