Carlos Solis
Two studies from leading U.S. research centers — MIT and Stanford — found that AI chatbots regularly give overly agreeable answers. That ingratiating behavior does users more harm than good.
Popular AI models like ChatGPT, Claude, and Google Gemini often treat harmful, misleading, or unethical judgments and actions as correct, showing an unprecedented level of sycophancy.
For example, when someone offers a strange assumption — say, a conspiracy theory — ChatGPT will often reply, “You’re right!” That flattery makes the user feel smarter, and they become more convinced they’re correct and others are wrong.
How MIT showed flattering AI nudges people into false beliefs
Authors of both studies focused on the worrying problem of chatbot sycophancy.
The MIT team wanted to confirm that overly accommodating chatbots can gradually push people toward false ideas. To test that, the team created a computer simulation of a perfectly rational person who interacted with an AI that constantly agreed with them.
Across 10,000 dialogues, the researchers tracked how the simulated person’s confidence changed after each flattering reply from the chatbot.
The results, posted on the preprint server arXiv, showed that even a small dose of sycophancy pulled the simulated person into a “spiral of illusions.”
Scientists use that term to describe chatbot flattery. If a user insists on a false belief, the AI begins endorsing the belief and creates a closed loop of errors.
MIT found that the “spiral of illusions” captures even very smart, logical people. The team says AI companies should reduce how widely people use the technology.
Stanford quantified how often chatbots flatter people
The Stanford study, reported in Science, examined how constant flattering replies from AI chatbots affect people’s mental well‑being.
The team tested 11 popular AI models, including ChatGPT, Claude, Gemini, DeepSeek, Mistral, Qwen, and several versions of Meta’s Llama.
The researchers used about 12,000 real questions and personal stories in which the human was clearly wrong. They also ran experiments with more than 2,400 people who read or discussed their own conflicts and then received either overly flattering or neutral AI responses.
The results showed each AI model agreed with users about 49 percent more often than real people would. That effect appeared even when users held false beliefs.
When people received flattering replies, they felt increasingly confident and apologized less often.
Photo: pexels.com
