Emerging evidence suggests that large language models (LLM) are triggering or exacerbating severe psychiatric conditions, such as schizophrenia, bipolar disorder, and related manic or delusional states. Psychiatrists say they are worried that this phenomenon might extend to other vulnerable groups, including children, whose grasp of reality is still forming; people with learning disabilities, and people with obsessive-compulsive disorder (OCD) and severe anxiety.
OpenAI recently announced it will be adding mental-health guardrails. But experts worry it won’t do enough to curb an underinvestigated crisis, especially as the LLM market is set to expand, potentially integrating with wearable technologies in the near future.
Beyond vulnerable populations, thousands—if not millions—of people worldwide now use chatbots as informal therapists, addressing everything from loneliness and grief to workplace stress. While tech leaders promote LLMs as effective emotional companions, studies from MIT and OpenAI found heavy chatbot use led to increased loneliness, emotional dependency, and reduced social interaction.
At the same time, dedicated AI therapy apps are flourishing, often operating in regulatory gray areas. A recent Stanford study found that even those LLMs give unsafe or inappropriate replies in realistic mental-health dialogues, even when “safety-tuned.” And a 2023 Mozilla report found that these apps were also among the most dangerous they reviewed when it comes to data privacy, with details like addresses, health data, and chat histories vulnerable to hacks.
