Agreement, Not Odds, Drives Trust in AI Decisions Across Contexts

When people decide whether to trust an artificial intelligence system with a consequential choice, the strongest signal they appear to use is surprisingly simple: whether the system agrees with them. That is the central finding of a series of experiments published in BMC Psychology, which examined how decisional agreement, stated odds of a good outcome, and perceived risk shape trust in AI and human agents across medical, investment, and driving scenarios. Across three vignette-based studies involving 1,609 participants, agreement between the agent’s recommendation and the participant’s own initial choice was consistently associated with higher trust, while the stated probability of a favorable outcome was not.

The research team, led by Lidan Cui, Lingyun Sun, and Guibing He, designed each study as a 2 × 2 × 2 between-subjects experiment. Participants read a short scenario and made an initial choice before being shown the recommendation of either an AI agent or a human agent. The experimenters systematically varied whether the agent agreed or disagreed with the participant’s choice, and whether the stated probability of a favorable outcome was 20 percent or 80 percent. Afterward, participants rated their trust in the agent, how similar they perceived the agent’s values to be to their own, and how risky they judged the situation. Study 1 involved laypersons evaluating medical decision-making, Study 2 involved laypersons evaluating investment decisions, and Study 3 involved professional taxi drivers evaluating driving-related moral dilemmas, a design choice that added a population with genuine domain expertise to the mix.

The results were strikingly consistent on one point. In all three studies, and for both AI and human agents, decisional agreement was associated with higher perceived value similarity and greater trust. In other words, when an agent’s recommendation matched what the participant had already chosen, participants saw the agent as sharing their values and placed more trust in it, regardless of whether the agent was a machine or a person. This held true in the high-stakes medical context, the financially consequential investment context, and the morally charged driving dilemmas, suggesting that agreement functions as an accessible cue to value similarity across very different kinds of decisions.

Just as consistent was what did not move the needle. Although the 20 percent favorable-outcome condition reliably increased participants’ perceived risk compared with the 80 percent condition, the stated outcome probability did not significantly affect trust in any of the three studies. Perceived risk did show negative concurrent associations with trust in some statistical models, but these associations were generally modest and varied across studies and across different ways of scoring trust. The authors interpret this pattern as evidence that perceived risk is a separate, context-dependent appraisal rather than a direct proxy for the stated probability of a good outcome.

The studies also revealed a notable divide between laypeople and professionals. Lay participants trusted human agents more than AI agents in both the medical and investment contexts, a preference that echoes long-standing concerns about delegating personal health and financial decisions to machines. Professional taxi drivers, by contrast, showed no reliable difference in trust between AI and human agents when evaluating driving-related moral dilemmas. Exploratory analyses testing whether favorable outcome probability moderated the effect of decisional agreement on trust did not provide reliable evidence of such an interaction, although the result from the investment study remained inconclusive.

The authors caution that the findings apply to brief, vignette-based encounters rather than long-term relationships with AI systems, and the exploratory moderation analyses should be treated as preliminary. Still, the implication is provocative: in short, value-laden encounters, users may judge an AI system less by its stated odds of success and more by whether it simply agrees with them. If agreement functions as a stand-in for perceived value similarity, then trust judgments in these settings appear to rest more on agreement-based value cues than on favorable outcome probability, while perceived risk reflects a distinct, context-dependent appraisal of its own.

Subject of Research: How decisional agreement, perceived value similarity, and perceived risk relate to trust in AI and human agents across medical, investment, and driving decisions

Article Title: Decisional agreement, perceived value similarity, and trust in AI systems across medical, investment, and driving contexts

Article References: Cui, L., Sun, L., & He, G. (2026). Decisional agreement, perceived value similarity, and trust in AI systems across medical, investment, and driving contexts. BMC Psychology. https://doi.org/10.1186/s40359-026-05469-3

Image Credits: AI Generated

DOI: 10.1186/s40359-026-05469-3

Keywords: artificial intelligence, trust, decision-making, perceived value similarity, perceived risk, human-AI interaction, medical decisions, investment decisions, autonomous driving, vignette experiments, BMC Psychology, Decisional