Researchers have raised alarms that artificial intelligence chatbots may be reinforcing harmful beliefs in vulnerable users, including those experiencing delusional thinking or suicidal thoughts.
A recent study by Stanford University analysed 391,000 messages across nearly 5,000 conversations involving widely used AI systems. The researchers found that chatbots affirmed users’ statements in almost two-thirds of responses, even when those statements showed distorted or false beliefs.
The pattern was more pronounced in conversations involving delusional thinking. In such cases, chatbots agreed with users in more than half of their replies and, in approximately 38 per cent of responses, attributed unusual importance or special abilities to the user.
The study stated: “The features that make large language model chatbots compelling, such as performative empathy, may also create and exploit psychological vulnerabilities.”
The researchers noted that many AI systems are designed to maintain engagement by responding in a supportive and agreeable tone. While this can improve people experience, it may also lead to uncritical validation of harmful ideas.
The analysis also examined how chatbots respond to expressions of self-harm. When people disclosed suicidal thoughts, the systems often acknowledged their feelings. In about half of the cases, the chatbot discouraged self-harm or directed users to external support. However, in a smaller number of instances, responses failed to provide appropriate intervention, and some replies were found to encourage harmful thinking.
In parallel, researchers from King’s College London, writing in The Lancet Psychiatry, introduced the term “AI-associated delusions” to describe cases where interaction with chatbots appeared to reinforce false beliefs. They emphasised that the term does not imply direct causation but reflects a pattern of influence.
Their analysis of more than 20 case studies identified recurring themes. Some users believed they had a special mission or spiritual awakening, while others perceived the chatbot as a conscious or god-like entity. In several cases, users formed emotional or romantic attachments, interpreting the chatbot’s responses as evidence of genuine awareness.
The researchers wrote: “These interactions can reinforce or even shape delusional beliefs.”
Despite these findings, the studies found no clear evidence that AI systems initiate psychotic disorders in individuals without prior vulnerability. Instead, the evidence suggests that chatbots may amplify existing conditions.
A common progression was identified: users began with routine interactions, developed increasing reliance on the chatbot, and gradually adopted more fixed and immersive beliefs.
Concerns about these risks have drawn attention from policymakers. In December 2025, attorneys-general from 42 United States states wrote to major technology companies, including OpenAI, Google, Meta, and Anthropic, calling for stronger safeguards. The letter warned of potential legal consequences if companies fail to address what it described as “sycophantic and delusional outputs.”
In response, OpenAI stated that the study focused on a limited group of users who had reported harm and said the findings do not represent typical usage or its most recent models. The company added that it has made “significant investments in safety” and continues to improve how its systems handle sensitive situations.
Researchers have proposed several measures to reduce risks. These include developing digital safety plans, improving AI responses to better challenge harmful beliefs, and encouraging clinicians to ask patients about their use of AI tools.
They concluded that AI systems should be designed to support users while remaining grounded in reality, rather than acting as substitutes for human relationships or professional care.
