Large language models capable of distinguishing between single and repeated gambles: Understanding and intervening in risky choice

Abstract: Risky choice (RC) is a common and important form of decision-making in daily life. Its theoretical development primarily follows two major approaches: normative theory and deThis paper comprises three studies: Study 1 employed GPT-3.5 and GPT-4 to simulate human responses to gambling decisions under nine probability conditions (with constant expected value), generating a total of 3,600 responses across both single and repeated gamble scenarios. Study 2 first constructed LLM-generated strategies through a three-stage process (decision rationale extraction, strategy generation, and quality evaluation). Human participants then completed decision-making tasks in two experiments: Experiment 1 replicated Sun et al.’s (2014) medical/financial scenarios (N = 349, Nmale = 174, Mage = 21.79) in a 2 (Context: Medical vs. Financial) × 2 (Application Frequency: Single vs. Repeated) within-subjects design, while Experiment 2 examined digital contexts with a 2 (Context: Content Creation vs. E-Commerce Marketing)×2 (Frequency: Single vs. Repeated) mixed design (context as between-subjects). Subsequently, DeepSeek-R1 performed the same tasks and generated strategy texts through the three-stage process. Finally, participants evaluated their acceptance of these LLM-generated strategies. Study 3 extended Study 2’s methodology to examine whether LLM-generated intervention texts could reverse participants’ classic choice preferences across single versus repeated gamble scenarios. Mirroring Study 2’s experimental contexts (Experiment 1: Medical vs. Financial, N = 460, Nmale = 205, Mage = 21.80; Experiment 2: Content Creation vs. E-Commerce Marketing, N = 240, Nmale = 106, Mage = 29.12), we presented strategically designed intervention texts during decision-making tasks to test their capacity to modify participants’ inherent risk preferences between single and repeated gamble conditions, thereby evaluating the persuasive efficacy of LLM-generated strategies on human decision biases.
Study 1 found that LLMs (GPT-3.5 and GPT-4) successfully replicated the typical human pattern of risk aversion in single-play scenarios and risk-seeking in repeated-play scenarios, although both models demonstrated an overall stronger tendency toward risk-seeking than humans. Study 2 found that human participants prefer low-EV certain options in single-play contexts and high-EV risky options in repeated-play contexts in both experiemnts. Participants also showed high agreement with the strategies generated by LLMs in different scenarios. Study 3 confirmed that LLM-generated intervention texts significantly influenced participants’ choice tendencies in all the four scenarios, with stronger intervention effects observed in single-play contexts. The LLMs’ intervention strategies were characterized by: reliance on expected value computations (normative) when promoting risky choices, and emphasis on certainty and robustness (deIn summary, this study demonstrates that: (1) LLMs can effectively simulate context-dependent human preferences in risk choice, particularly the shift from risk aversion in single plays to risk-seeking in repeated plays; (2) LLMs can distinguish between the logic underlying single and repeated gambles and apply both normative and descriptive reasoning accordingly to externalize decision strategies; and (3) decision strategies extracted from LLM-generated reasoning can be used to construct effective intervention texts that alter human preferences in classic risk decision tasks, thereby validating the feasibility and effectiveness of an LLM-based cognitive intervention pathway. This study offers a new technological paradigm for AI-assisted decision intervention and expands the application boundary of LLMs in modeling and regulating human cognitive processes.