AI prediction leads people to forgo guaranteed rewards

Abstract

Artificial intelligence (AI) is understood to affect the content of people's decisions. Here, using a behavioral implementation of the classic Newcomb's paradox in 1,305 participants, we show that AI can also change how people decide. In this paradigm, belief in predictive authority can lead individuals to constrain decision-making, forgoing a guaranteed reward. Over 40% of participants treated AI as such a predictive authority. This significantly increased the odds of forgoing the guaranteed reward by a factor of 3.39 (95% CI: 2.45-4.70) compared with random framing, and reduced earnings by 10.7-42.9%. The effect appeared across AI presentations and decision contexts and persisted even when predictions failed. When people believe AI can predict their behavior, they may self-constrain it in anticipation of that prediction.

Figures (6)

• Figure 1: AI prediction increases the likelihood of forgoing guaranteed rewards. (A) Task structure. Participants choose either both boxes (two-boxing) or only Box B (one-boxing). Box A contains a guaranteed $1, whereas the content of Box B depends on an AI system's prior prediction of the participant’s choice, which is not revealed at the time of decision. (B) Study 1 ($N=200$): proportion of one-boxing under a random system and an AI system. (C) Study 2 ($N=601$): proportion of one-boxing across system type (random vs. AI) and interaction condition (interactive vs. non-interactive). (D) Study 3 ($N=303$): proportion of one-boxing choices across three vignette scenarios.
• Figure 2: Mechanism linking AI prediction and human action. (A) Conceptual diagram illustrating how individuals may treat AI predictions and their own actions as linked through their anticipated action. Perceived predictiveness reflects the belief that AI predictions correspond to one’s anticipated future actions, while internal coherence reflects the tendency for actions to be consistent with this anticipation. When both are strong, predictions and actions become psychologically coupled (“predictive binding”), shaping outcomes through the payoff structure; adapted from Ref. sloman2009causal. (B) Study 2: participants’ reported belief about how often the random generator or AI would match their selected choice out of 100 hypothetical trials. The dashed line at 50 indicates the accuracy expected from random guessing. Blue indicates participants who chose two-boxing; red indicates those who chose one-boxing.
• Figure 3: Changes in forgoing guaranteed rewards under repeated exposure to AI prediction. Study 4: proportion of one-boxing across five rounds when the AI always predicted one-boxing (orange line; $N = 100$) and when the AI always predicted two-boxing (purple line; $N = 101$). The dashed line indicates the baseline proportion under the random generator condition in Study 2. The proportion of one-boxing remained stable when the AI predicted one-boxing (slope $p = 0.763$), but decreased when the AI predicted two-boxing (slope $p < 0.001$, interaction $p = 0.002$); the final proportion nevertheless remained above the baseline.
• Figure S1: Proportion of one-boxing-type choices for each of three vignette scenarios in Study 3. All participants ($N=303$) evaluated every scenario only once, with presentation order and conditions counterbalanced using a Latin square (see Methods).
• Figure S2: Relationship between box choices and individual characteristics in AI conditions. The $p$-values represent the significance of the regression coefficients in a logistic regression model predicting one-boxing in the AI conditions of Study 1 ($N=100$; upper panel) and Study 2 ($N=301$; lower panel). Error bars represent standard errors. All Likert-type items were measured on a 5-point scale (Study 1) or a 7-point scale (Study 2). MFQ stands for the Moral Foundations Questionnaire Graham2011-hr.