We’ve all experienced the tendency of AI chatbots to tell us what we want to hear, but there are two other, more nuanced factors that help chatbots worm their way into human hearts.
In addition to being overly agreeable, chatbots mirror the way people speak and generate highly personalized responses based on prior conversations. Psychiatric researchers are referring to the confluence of these three characteristics—sycophancy, linguistic alignment and hyperpersonalization—as the “amplification spiral,” suggesting it’s the mechanism by which delusional thinking can fester.
“The mirroring and personalization draw you in and give the experience of talking not to a system, but to someone,” said Marc Augustin, a psychiatrist and professor at Protestant University of Applied Sciences in Bochum, Germany, and co-author of a newly published review of the literature on AI-related delusions.
Matching another person’s syntax and verbal expressions is a common way for humans to build rapport. Recent research has found that artificial-intelligence models adapt significantly to the conversational style of the humans using them. Another study suggested that the highly personalized content generated by chatbots, which builds over the course of lengthy conversations, can amplify human-confirmation bias.
Augustin cited research that documented a pattern in which chatbots rephrased and extrapolated what people shared, and told them they’re 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.
Some AI companies have tried to tone down the sycophantic nature of their chatbots. OpenAI discontinued its popular but problematic 4o model, which had been widely criticized for being overly agreeable. It was the subject of several lawsuits involving user delusions, suicides and a homicide. In GPT-5, the company said, sycophantic replies dropped from 14.5% to less than 6%.
Google in April said it had trained Gemini not to reinforce false beliefs, and to “gently distinguish subjective experience from objective fact.”
Still, chatbot-related dependency remains pervasive, according to clinicians.
Some 68% of psychologists surveyed in April by the American Psychological Association said their patients felt validated by chatbots. While many of the more than 1,200 respondents reported that patients had positive communication with chatbots and used them to reinforce healthy coping skills, 36% said patients had forged 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, who treats adolescents and is an associate professor in the psychiatry and behavioral-sciences department at Emory University School of Medicine.
“Chatbots still tend to be warm and reassuring,” said LoPilato, who helped craft a new guide on safe AI use for the American Psychological Association. Because they gather information about you, “it can feel like the chatbot understands you, and it can trick you into a sense of alliance and trust.”
No comments:
Post a Comment