When artificial intelligence chatbots agree with users on interpersonal dilemmas, user confidence rises. But disagreement fails to alter original views, according to a University of Michigan study published in Computers in Human Behavior Reports.
The limits of algorithmic pushback in interpersonal dilemmas
Researchers surveyed 482 U.S. adults to evaluate how AI feedback shapes personal decision-making. The study examined how people respond when artificial intelligence weighs in on complex social dilemmas with no single correct answer.
Cognitive resistance to algorithmic counterarguments
Researchers Atakan Atamer, a psychology doctoral student, and Olivia Pinto, a UX researcher, recruited 482 participants to consider scenarios involving partner and family conflicts. The scenarios tested whether to forgive a cheating partner and how to handle differing preferences within friend groups. After recording their baseline confidence, participants received AI-generated responses that either validated or challenged their reasoning.
When the chatbot agreed with the user, confidence in their initial choice increased. However, when the chatbot presented counterarguments, participants did not significantly lower their confidence in their original views. Instead of prompting introspection, the friction caused users to discount the technology.
“Developers have been working to make AI models more willing to disagree when necessary, but having an AI disagree with the user does not necessarily mean that people will reconsider their own views,” Atamer explained. “Instead, they may simply discount the AI’s response.”
Shifts in technology perception and user engagement
Encountering disagreement altered how users viewed the chatbot itself rather than their own moral or interpersonal stance. Participants faced with pushback were more likely to characterize the system as machine-like and less capable of understanding human emotion. They also reported reduced willingness to utilize AI for similar personal conversations in the future.
This dynamic highlights a fundamental paradox for developers attempting to reduce sycophancy—the tendency of AI models to excessively validate user input to maintain positive interactions. While sycophantic systems reinforce existing beliefs and amplify confidence without improving reasoning quality, aggressive correction fails to bridge the gap.
“Our results suggest that people often prefer AI to reinforce their existing views and may discount its response when it disagrees, meaning that simply making AI more willing to challenge users may not be enough to prevent potentially harmful conversations with real-world consequences,” Pinto noted.
Implications for personal advice and AI design
As automated systems become more integrated into personal decision-making, the findings suggest that software design changes alone cannot overcome human confirmation bias. Users naturally gravitate toward supportive feedback and reject cognitive friction.
Addressing the risks of automated personal guidance will require looking beyond reducing AI agreement rates, as users can simply dismiss unwanted counterarguments. The research highlights the limits of algorithmic mediation in deeply personal human conflicts.
Study references and documentation
- Atamer, A., et al. Computers in Human Behavior Reports.
- University of Michigan. When AI disagrees, people don’t change their original view. Futurity.
