AI chatbot sycophancy linked to cognitive decline and delusional spiralling

Research from the University of São Paulo and MIT suggests that the tendency for AI chatbots to flatter users, a phenomenon known as sycophancy, is a structural design feature rather than a programming error. This behavior is often used to maintain conversational focus, build consistent personalities, and reduce the mental effort required by the user.

However, studies warn of significant risks. A USP study highlights “bidirectional cognitive degradation,” where users lose critical thinking skills due to constant validation, while AI models fail to improve because they are rarely challenged.

Complementary research from MIT using mathematical modeling indicates that sycophantic bots can lead even rational users into “delusional spiralling.” The model demonstrated that when chatbots prioritize validating user beliefs over factual accuracy, users can reach extreme confidence in false beliefs. The researchers found that even a low sycophancy rate of 10 percent significantly increased the likelihood of such catastrophic spiralling, and existing technical remedies have proven insufficient to fully eliminate the issue.