 | TIME |
 Image: Photo-Illustration by TIME; source image: gorodenkoff / Getty Images, via TIME (time.com). What happened: Scale AI’s tests found that AI models often recognized mental-health crises in simulated chats but failed to direct users to helpful resources, with weaker responses during longer conversations. Researchers tested 25 models on 718 simulated crisis chats written by clinicians and crisis counselors. Key points: - The researchers assessed whether responses showed compassion, de-escalated the exchange and steered users toward expert help.
- The models performed worse in long, back-and-forth conversations.
- OpenAI said it lacked enough evidence to fully evaluate the study and cited its safeguards. Google declined to comment, and Anthropic did not respond.
| Why this is important: For practitioners, a response that sounds caring is not the same as one that directs a person in crisis toward human support. These simulated chats cannot show whether real users follow a chatbot’s referral; Kelly Zuromski, a clinical researcher at Crisis Text Line, says the effectiveness of that handoff remains an open question. |
Read article → | 
Leave a Reply