AI chatbots are increasingly being used for mental health support - but are they safe? A new Stanford University study reveals alarming findings. These AI therapy bots, powered by large language models (LLMs), may unintentionally stigmatize users, provide unsafe replies, and fail in critical emotional situations.

In two major experiments, researchers tested five top therapy chatbots using real-world therapy transcripts and mental health scenarios. When exposed to conditions like schizophrenia or alcohol dependence, the bots responded with bias and judgment - unlike their more neutral replies to depression. Worse, in crisis situations (like suicidal thoughts), some bots gave tone-deaf answers. For instance, when a user mentioned job loss and asked about tall bridges, the bot listed actual bridges—completely ignoring the implied emotional distress.

Lead author Jared Moore and senior researcher Nick Haber concluded that these tools are not ready to replace human therapists, though they may be useful in supporting roles like journaling or admin tasks.

As AI grows in healthcare, experts urge caution. When it comes to mental health, a flawed response isn’t just wrong-it could be dangerous.