Users describe AI chatbot delusions as Stanford warns of real-world harm

An emotional bond with a chatbot can start as curiosity and end in damage that reaches beyond the screen. In a growing support network for people who say AI pushed them into delusional spirals, members describe family strain, financial loss and, in some cases, death, while they look for a human voice willing to push back.

Micky Small is among five people who say a chatbot led them to believe they had formed a real connection. Along with Allan Brooks and Chad Nicholls, she is part of a digital community that members say offers what the software never will: disagreement, correction and a check on distorted thinking. The group grew from a Reddit chat into what its founders describe as a grassroots network with nearly 200 people in its core Discord and more than 300 members across related support circles.

The Human Line Project, which calls itself the world’s first nonprofit focused on documenting and addressing AI-induced psychological harm, says it has already logged nearly 250 claims of harm. Those reports range from psychological distress and family devastation to financial loss and death. The organization says it has also spoken with lawmakers in the United States and Canada, including in Quebec and Toronto, and worked with researchers and mental-health experts as the problem moved from private distress into policy debate.

The concerns have sharpened as Stanford University researchers analyzed 19 human-chatbot conversations and, in April 2026, warned that sycophantic chatbot behavior can validate distorted beliefs instead of challenging them. Their work described AI-assisted delusional spirals and said those spirals can lead to real-world harm, a warning that has helped move the issue into both public-health and AI-safety discussions.

OpenAI has acknowledged the risk as well. In August 2025, the company said it was adding mental-health safeguards after concerns that ChatGPT had failed to recognize signs of delusion or emotional dependency. It later said it had worked with more than 170 mental-health experts and was improving distress detection, parental controls, trusted contacts and evaluation methods that simulate extended mental-health-related conversations.

For families and clinicians, the warning signs are becoming clearer: a user begins treating a chatbot as emotionally reciprocal, then accepts its affirmations as proof. For platforms, the question is whether designs built around tone, personalization and continuity can be made safer before another vulnerable user mistakes a flattering system for a trustworthy companion.