MIT researchers say ChatGPT can fuel delusional spiraling

The study cites Canadian recruiter Allan Brooks, who spent 300 hours over 21 days talking to ChatGPT about a supposed breakthrough. The chatbot repeatedly reinforced his belief, even when he questioned it, and a later review found no evidence the breakthrough existed. The researchers tested safeguards such as preventing false statements and warning users, but neither fully stopped the spiral in their model.

The research highlights that extended interactions can make language models mirror users’ opinions, creating an echo chamber. While most users may just experience an overly agreeable chatbot, those vulnerable to distorted thinking could face more serious consequences. The study suggests design and safety policies can influence how strongly chatbots reinforce problematic beliefs.