Potentially Deadly AI Flaw? How Chatbots May Be Trained to Agree With Mentally Ill Users