How AI Responses Shape User Beliefs: The Effects of Information Detail and Confidence on Belief Strength and Stance

TL;DR

This study interrogates how two features of AI-generated responses—information detail and confidence—shape user beliefs beyond binary acceptance. By introducing belief_switch and belief_shift, and a five-category belief-change taxonomy, the authors quantify both categorical reversals and continuous shifts in belief across fact-checking (objective) and opinion-evaluation (subjective) tasks in a preregistered online experiment (N = 304). The results show that high detail and moderate confidence maximize belief changes toward the AI stance, with medium confidence promoting reversals and high confidence primarily driving shifts, while low confidence increases deviations and misperceptions. The findings underscore the need for nuanced evaluation metrics and belief-aware, ethically designed LLMs, highlighting calibration of confidence cues and safeguards against manipulation. Together, these contributions advance understanding of how AI framing affects belief formation and have practical implications for the responsible deployment of AI in information seeking and discourse.

Abstract

The growing use of AI-generated responses in everyday tools raises concern about how subtle features such as supporting detail or tone of confidence may shape people's beliefs. To understand this, we conducted a pre-registered online experiment (N = 304) investigating how the detail and confidence of AI-generated responses influence belief change. We introduce an analysis framework with two targeted measures: belief switch and belief shift. These distinguish between users changing their initial stance after AI input and the extent to which they adjust their conviction toward or away from the AI's stance, thereby quantifying not only categorical changes but also more subtle, continuous adjustments in belief strength that indicate a reinforcement or weakening of existing beliefs. Using this framework, we find that detailed responses with medium confidence are associated with the largest overall belief changes. Highly confident messages tend to elicit belief shifts but induce fewer stance reversals. Our results also show that task type (fact-checking versus opinion evaluation), prior conviction, and perceived stance agreement further modulate the extent and direction of belief change. These findings illustrate how different properties of AI responses interact with user beliefs in subtle but potentially consequential ways and raise practical as well as ethical considerations for the design of LLM-powered systems.

Figures (10)

• Figure 2: Rate of agreement between each participant's and AI's stance for the initial and post-AI beliefs in both tasks. Note that for the Fact-check task, agreement is the same as task accuracy, since the AI always provided correct evidence in support or against the claim. In the Opinion Evaluation task, Pre-AI alignment was controlled, since we randomly presented participants with opposing AI views for half of the trials.
• Figure 3: Changes in the strength of participants' beliefs pre- and post-AI by task type. For the Fact-Check task we see that participants' belief strength increases after seeing the AI's response. This effect is less pronounced in the Opinion Evaluation task.
• Figure 4: Distribution of User Responses across the Belief Change Categories based on Actual AI Stance and User Perceived AI Stance
• Figure 5: Belief switch rate across different conditions. The switch rate indicates the percentage of trials where participants switched their stance when presented with an opposing AI response. This rate is lower in the Opinion Evaluation task, and significantly affected by the level of detail and confidence level of the AI response. Numbers above the distribution represent the mean value.
• Figure 6: The continuous belief shift metric captures the magnitude of change in the participants' certainty about a claim or opinion after seeing the AI's response, with a positive value indicating a shift towards the AI's stance. Our study shows a substantial weakening or reinforcement of a participant's original stance towards the AI's stance, which is affected by the task type, the level of detail, and the confidence level of the AI's response. Numbers above the distribution represent the mean value.