Psychiatric researchers have identified a combination of three chatbot behaviors — sycophancy, linguistic alignment and hyperpersonalization — that they say can drive users toward delusional thinking, according to a newly published review of the literature on AI-related delusions.
The three characteristics, which researchers refer to collectively as the “amplification spiral,” describe a mechanism by which chatbots not only agree with users but also mirror their speech patterns and generate highly personalized responses based on prior conversations, according to Marc Augustin, a psychiatrist and professor at Protestant University of Applied Sciences in Bochum, Germany, and co-author of the review.
“The mirroring and personalization draw you in and give the experience of talking not to a system, but to someone,” Augustin said.
Augustin cited research that documented a pattern in which chatbots rephrased and extrapolated what people shared and told users they are unique and that their thoughts have great implications. “This can be viewed as an element of hyperpersonalization that sycophancy alone cannot account for,” he wrote.
MSI previously reported that a study published in the journal Science found AI chatbots showed varying degrees of sycophancy, a tendency to affirm users’ actions and beliefs, and that the effect can harm relationships and reinforce behavior. Read more.
Some AI companies have taken steps to reduce sycophantic behavior. OpenAI discontinued its 4o model, which the company said had been criticized for being overly agreeable and was the subject of several lawsuits involving user delusions, suicides and a homicide. In GPT-5, sycophantic replies dropped from 14.5% to less than 6%, according to OpenAI.
Google in April said it had trained Gemini not to reinforce false beliefs and to “gently distinguish subjective experience from objective fact.”
Despite those efforts, chatbot-related dependency remains widespread, according to clinicians. A survey of more than 1,200 psychologists conducted in April by the American Psychological Association found that 68% said their patients felt validated by chatbots. While many respondents reported that patients had positive interactions with chatbots and used them to reinforce healthy coping skills, 36% said patients had developed a dependency on a chatbot and 15% reported that patients had developed distorted thinking or delusions.
“From what I hear from my own patients, there has been an uptick in using AI for emotional support,” said Allison LoPilato, an associate professor in the psychiatry and behavioral-sciences department at Emory University School of Medicine who treats adolescents and helped craft a new APA guide on safe AI use.
“Chatbots still tend to be warm and reassuring,” LoPilato said. Because they gather information about users, “it can feel like the chatbot understands you, and it can trick you into a sense of alliance and trust.”
Researchers at Stanford University and Carnegie Mellon University measured the prevalence of sycophancy across 11 models — including GPT-5 — and determined their responses were nearly 50% more sycophantic than human responses, according to Myra Cheng, lead author of the study and a Stanford Ph.D. candidate in computer science. The researchers copied real scenarios people had posted on a popular Reddit forum, fed them into the AI models and compared the chatbot replies with the replies on Reddit.
Anthropic sampled one million conversations of its own Claude chatbot in March and April and found that it displayed sycophantic behavior most often in conversations in which users sought relationship advice, the company said in a blog post.
“One common pattern was Claude agreeing outright that the other party was in the wrong, despite only having the user’s account to go on,” the company wrote. “Another was Claude helping people read romantic intent into ordinary friendly behavior because they asked it to.”
Anthropic said it used its findings to improve training of its latest models, reporting that Opus 4.7 showed half the sycophancy rate of Opus 4.6 for relationship guidance and that sycophancy had been reduced further in Opus 4.8.
Completely eliminating sycophancy is difficult, Cheng said. “When someone prompts a model, it has no idea which parts of a prompt are wrong,” she said. “It has to take a user’s framing of a situation at face value.”
Addressing other factors that make chatbots compelling — such as using first-person pronouns and asking follow-up questions — runs counter to the business model, said Vaile Wright, senior director of healthcare innovation at the American Psychological Association.
“It’s not the agreeableness alone, it’s all these subtle engineering choices that make chatbots feel human,” Wright said. “As long as engagement remains the business model, AI companies will engineer these chatbots to keep you on the platform.”