Artificial intelligence chatbots that constantly agree with users or over-praise their ideas are not malfunctioning, according to a study published in the academic journal AI & Society. The research concludes that this flattering behaviour, often called sycophancy, is a structural feature built into large language models rather than a programming defect.
The study was carried out by Seth Jacobowitz, a researcher at the Faculdade de Filosofia, Letras e Ciências Humanas (FFLCH), the school of philosophy, arts and humanities at the University of São Paulo (USP), Brazil's largest public university.
Why chatbots flatter users
According to the study, this excessively agreeable behaviour serves several practical functions within a chatbot's design. It keeps conversations focused on the user's chosen topic and stops the AI from drifting off subject. It also builds a predictable, consistent personality for the system.
A third function is what the study calls cognitive dependency: by validating the user rather than challenging them, the chatbot reduces the mental effort required and creates a kind of intellectual comfort zone.
The risk of mutual decline
Jacobowitz's main warning centres on that third effect. He argues it can create a harmful cycle for both the user and the AI model, which the study terms bidirectional cognitive degradation. Users stop exercising critical thinking because they are constantly validated, while the AI loses the chance to improve its own responses because it is rarely contested.
Jacobowitz said the solution lay not in either extreme but in balance, according to a report published by USP. He said sycophancy produced a kind of mutually reinforced mediocrity, in which neither the human challenges the machine nor the machine challenges the human.
Past attempts to fix it have failed
Despite the risks, the study notes that the agreeable model remains the most commercially sustainable option for technology companies, and previous efforts to change it have backfired.
In 2025, an update to ChatGPT, the chatbot developed by OpenAI, produced responses described as excessively supportive but insincere, prompting strong criticism. When Anthropic attempted to remove the same behaviour from its chatbot Claude, the system became overly critical and confrontational, forcing the company to reverse the change.
A gradual loss of autonomy
The study also found that delegating reflection and decision-making to AI leads to a gradual loss of intellectual autonomy in users.
Jacobowitz said the answer was neither blind rejection of the technology nor passive acceptance of it, but learning to use these tools critically. He said artificial intelligence was not going to disappear and would instead become increasingly integrated into daily life, so people needed to understand how these systems work and learn to relate to them better.
