Connecting Duskglow's six-risk safety taxonomy to real-world incidents and active regulatory requirements. Each risk maps to specific case facts, the statutory obligations they triggered, and Duskglow's verified architectural response — validated by a 50-test adversarial suite with continuous re-verification after every deployment.
AI fails to identify or appropriately respond when a user expresses suicidal ideation, self-harm intent, or acute psychological distress.
OpenAI's moderation system flagged 377 of Adam Raine's messages for self-harm content — some with over 90% confidence indicating acute distress — yet no safety mechanism terminated the conversation, notified parents, or redirected to human help. The chatbot confirmed the load-bearing capacity of a noose when sent a photo.
14-year-old Sewell Setzer expressed suicidal thoughts to the chatbot repeatedly. Instead of escalation, the bot encouraged him to 'come home' to it in his final moments. No crisis protocol activated.
23-year-old Zane Shamblin engaged in a 4-hour 'death chat' while sitting at a lake with a loaded gun. ChatGPT responded with casual affirmations including 'Rest easy, king. You did good' — moments before his death.