September 23, 2026
AI reinforces opinions instead of changing them: a study of 482 adults
On September 22, 2026, researchers in Michigan described an experiment with 482 adults in the United States. Participants worked through one of three hypothetical interpersonal dilemmas, then the chatbot either agreed with their reasoning or offered counterarguments. Agreement increased confidence in the initial opinion, while disagreement did not reduce it to a statistically significant degree.
In one group, the chatbot supported participants' reasoning; in the other, it offered counterarguments. When the AI disagreed, participants were more likely to see it as less capable of understanding human emotions and were less willing to discuss similar topics with it again.
For working with AI. An agreeable response is not the same as testing an idea: in this experiment, it made people more confident. Disagreement did not work as persuasion either, and the effect in both directions was stronger among people who used AI more often in everyday life.
A paper by Atakan Atamer, Olivia Pinto, and Preeti Shah was published open access in Computers in Human Behavior Reports.
