2026-09-22 ミシガン大学
Split-screen illustration showing the same woman interacting with an AI chatbot. Image credit: Nicole Smith, made with ChatGPT
＜関連情報＞
- https://news.umich.edu/when-ai-disagrees-people-change-how-they-perceive-the-technologynot-their-original-view/
- https://www.sciencedirect.com/science/article/pii/S245195882600326X
味方からアルゴリズムへ：意見の相違がAIとの関わり方をどのように形作るか From ally to algorithm: How disagreement shapes users’ engagement with AI
Atakan Atamer, Olivia Pinto, Priti Shah
Computers in Human Behavior Reports Available online: 12 August 2026
DOI:https://doi.org/10.1016/j.chbr.2026.101252
Highlights
- Participants were randomly assigned to discuss social dilemmas with an AI chatbot that either agreed or disagreed with their reasoning.
- Participants perceived the AI as less capable of understanding human emotions and more machine-like when it disagreed with them.
- AI input had asymmetric effects on participants’ confidence: agreement increased confidence in their initial opinion, whereas disagreement did not significantly reduce confidence.
- Participants reported greater intentions to use AI again for similar conversations when it agreed with them rather than disagreed.
- These effects were stronger among participants who used AI more frequently in their daily lives. Overall, our findings suggest that even when models are less sycophantic, people may still dismiss disagreement through asymmetric confidence update and reduced willingness to reuse the AI.
Abstract
Advances in AI language capabilities have led people to use it for conversations about relationships, social dilemmas, and moral questions. However, AI chatbots are sycophantic; they tend to affirm what the user says, exacerbating confirmation bias and leading to overconfidence or distorted perceptions of reality. Based on these concerns, developers started addressing sycophancy problems, creating models more likely to disagree when necessary. Here, we investigated what happens when AI models disagree with users in conversations. Participants were presented with one of three hypothetical dilemmas in which they formed an opinion and explained their reasoning to an AI model. They were randomly assigned to the agree condition, in which the AI model was instructed to support the user’s reasoning, or the disagree condition, in which the AI model was instructed to provide refutations. Finally, participants rated AI’s capacity to understand human emotions, confidence in their decisions, and willingness to use AI again. Participants rated AI as less capable of understanding human emotions and as more machine-like in the disagree condition compared to the agree condition. Participants also weighed the AI’s opinion asymmetrically, increasing confidence when the model agreed but not decreasing it when it disagreed. Finally, they reported greater willingness to use AI to discuss interpersonal issues when it agreed with them. These findings suggest that people’s perception of AI depends on whether it agrees with them, and that they may discount disagreement by weighing its advice less. Hence, creating human–AI interactions may require more than reducing sycophancy.
