Passing the Buck to AI: How Individuals' Decision-Making Patterns Affect Reliance on AI

TL;DR

The study investigates how individuals' decision-making patterns influence their use of AI in decision tasks. Using a preregistered online experiment with 810 participants evaluating nutrition statements and optional AI decisions/explanations from ChatGPT, it finds that buckpassing increases the likelihood of seeking AI input and reporting reliance, while vigilance increases time spent evaluating AI explanations. Hypervigilance shows no consistent effect, and overall performance depends on AI accuracy, with misstatements reducing accuracy when consulted. The findings highlight the need to tailor AI explanations and interfaces to users' cognitive styles, and they provide a publicly available dataset linking decision-making patterns to AI interactions, with implications for designing safer, more effective AI-assisted decision systems.

Abstract

Psychological research has identified different patterns individuals have while making decisions, such as vigilance (making decisions after thorough information gathering), hypervigilance (rushed and anxious decision-making), and buckpassing (deferring decisions to others). We examine whether these decision-making patterns shape peoples' likelihood of seeking out or relying on AI. In an online experiment with 810 participants tasked with distinguishing food facts from myths, we found that a higher buckpassing tendency was positively correlated with both seeking out and relying on AI suggestions, while being negatively correlated with the time spent reading AI explanations. In contrast, the higher a participant tended towards vigilance, the more carefully they scrutinized the AI's information, as indicated by an increased time spent looking through the AI's explanations. These findings suggest that a person's decision-making pattern plays a significant role in their adoption and reliance on AI, which provides a new understanding of individual differences in AI-assisted decision-making.

Figures (10)

• Figure 1: The study interface and workflow for each statement. Participants first read one statement at a time. They could then decide whether the statement is a fact or myth (a). If participants decided at this point, they could move on to the next statement. Alternatively, participants could choose to reveal the AI's decision (1) showing ChatGPT's decision (b). Participants could then decide to choose fact or myth or use the AI decision. They could also choose to see AI's explanation to reveal the AI explanation before making their decision (2).
• Figure 2: Overview of Frequency of See AI Decisions with Different Levels of Buckpassing: Participants with varying levels of buckpassing show differences in how often they view AI decisions. "Low" and "High" buckpassing refer to scores that are more than one standard deviation below and above the average, respectively. "Average" buckpassing refers to scores that are within one standard deviation of the average. Each bar represents the average frequency, with error bars indicating the confidence intervals. On average, those who score low in buckpassing view AI decisions 25% of the time, while those who score high view them 33% of the time.
• Figure 3: Perceived AI Reliance Variations Across Different Levels of Buckpassing: individuals who scored high (one standard deviation above average among participants) in buckpassing reported a higher reliance on AI compared to those who scored low (one standard deviation below average) in buckpassing.
• Figure 4: Examining the demographic distribution of buckpassing tendency: participants who have a high school education or below score higher in buckpassing (23% in high buckpassing level); participants who are aged under 18 and in their early 20s also tend to score higher in buckpassing. The y-axis is the proportion of participants in each buckpassing group.
• Figure 5: Example of result page in the study presented to participants: we included an example of our result page regarding how we debriefed participants about their performance in the nutrition evaluation task. Specifically, we included the statements where they submitted inaccurate answers and statements where AI suggestions were inaccurate.