Context Over Language: Persona-Driven Intertemporal Choice in Large Language Models

Large language models (LLMs) are increasingly used as experimental subjects in behavioral and economic research. However, it remains unclear whether their decisions reflect stable preferences or context-dependent responses. This study examines intertemporal choice behavior in LLMs within a controlled experimental framework. We implement a fully factorial design across four models (ChatGPT, Gemini, Copilot, and Claude), varying persona (financially stressed vs. disciplined), delay (1 vs. 6 months), reward magnitude (650–900 $), and prompt framing (direct vs. human-like). Each condition is repeated to approximate response distributions. Results show that LLM behavior is structured but not preference-stable. Decisions are primarily driven by persona conditioning: financially disciplined agents choose delayed rewards significantly more often than stressed agents. Delay and reward magnitude affect choices in line with temporal discounting, while prompt framing has negligible effects. We interpret these findings by conceptualizing LLMs as context-dependent behavioral simulators rather than agents with stable preferences. The contribution is not to establish human equivalence, but to identify the internal consistency and structure of LLM responses under controlled variation. Overall, the results highlight the importance of experimental design and suggest that LLM outputs should be understood as conditional simulations rather than stable expressions of preference.