aceModerate evidence
More
People trust AI less when it disagrees, and shrug off its pushback, an experiment finds
Study title: From ally to algorithm: How disagreement shapes users’ engagement with AI
- Evidence type
- Randomized controlled trial
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-2571cite or check it
Study title: From ally to algorithm: How disagreement shapes users’ engagement with AI
A randomized experiment published in Computers in Human Behavior Reports tested what happens when an AI model disagrees with people discussing a dilemma. Participants formed an opinion on one of three hypothetical dilemmas and explained their reasoning to an AI model that was instructed either to support it or to refute it.
Read the study 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…
The problem
In a randomized experiment, participants rated a disagreeing AI model as less able to understand human emotions and more machine-like, did not lower their confidence when it disagreed, and were less willing to use it again.
In brief
Those who got pushback rated the AI as less capable of understanding human emotions and as more machine-like than those who got agreement. They also said they were less willing to use AI again to discuss issues. Confidence rose when the model agreed but did not fall when it disagreed.
The authors conclude that building human–AI interactions may require more than reducing sycophancy, because people may discount disagreement and weigh the advice less. The dilemmas were hypothetical, and the AI's agreement or disagreement was assigned by instruction.
Main points
- Participants were randomly assigned to an agree condition, in which the AI supported their reasoning, or a disagree condition, in which it offered refutations.
- Each participant formed an opinion on one of three hypothetical dilemmas and explained their reasoning to the AI model.
- In the disagree condition, participants rated the AI as less capable of understanding human emotions and as more machine-like.
- Confidence in decisions rose when the AI agreed but did not fall when it disagreed.
- Participants reported greater willingness to use AI to discuss issues when it agreed with them.
- The research was motivated by concerns that sycophantic chatbots affirm users, worsening confirmation bias and overconfidence.
The full story
Developers have begun trying to curb sycophancy, building models more likely to disagree when necessary, because chatbots tend to affirm what users say. People increasingly turn to AI for conversations about relationships, social dilemmas and moral questions.
The experiment measured three outcomes after the conversation: participants' rating of the AI's capacity to understand human emotions, their confidence in their own decisions, and their willingness to use AI again.
Results showed an asymmetry: agreement raised confidence, but disagreement did not lower it. The authors read this as people discounting disagreement by weighing the AI's advice less, and argue that reducing sycophancy alone will not make human–AI interaction work well.
Sources
1 on file- Computers in Human Behavior Reportsdoi:10.1016/j.chbr.2026.101252
Talk it over
Who carries this cost, and what would have to change to prevent it?
People trust AI less when it disagrees, and shrug off its pushback, an experiment findshttps://truvace.com/item/from-ally-to-algorithm-how-disagreement-shapes-users-engagement-with-ai
Who carries this cost, and what would have to change to prevent it?
The problem In a randomized experiment, participants rated a disagreeing AI model as less able to understand human emotions and more machine-like, did not lower their confidence when it disagreed, and were less willing to use it again.
Answer in the threadHow do we know? What can this kind of evidence show, and what can’t it?
The evidence This rests on a randomized controlled trial, from Computers in Human Behavior Reports.
Answer in the threadWhat does it leave open? What else would you want to know before acting on it?
The record’s own caveat Participants discussed hypothetical dilemmas with an AI instructed to agree or disagree, so the findings may not carry over to real personal decisions or to models that choose when to disagree.
Answer in the threadHave you seen AI used this way where you study, work or live? Did it go the same way?
Answer in the thread
The questions are the same on every record. The details under them are taken from this record as written.
What readers say