Rethinking Prospect Theory for LLMs: Revealing the Instability of Decision-Making under Epistemic Uncertainty

Researchers challenge the applicability of Prospect Theory to Large Language Models, finding that PT parameters are unstable when models encounter epistemic uncertainty markers like "likely" or "probably." The study warns against deploying PT-based frameworks in real-world applications where linguistic ambiguity is common, raising critical questions about LLM decision-making reliability.

This research exposes a fundamental gap between how behavioral economics models human decision-making and how those same models perform when applied to LLMs. The three-stage experimental design methodically tests Prospect Theory's robustness by first establishing PT parameters through economics questions, then introducing epistemic markers—common linguistic expressions of uncertainty—to measure parameter stability. The findings reveal significant inconsistencies across different models, suggesting that LLMs don't reliably conform to PT predictions under epistemic uncertainty.

The implications extend beyond academic interest. As LLMs increasingly power financial decision-making systems, trading algorithms, and risk assessment tools, their decision-making stability becomes commercially critical. Prospect Theory has been the theoretical foundation for understanding how to optimize LLM behavior in high-stakes domains. If PT parameters shift substantially based on subtle linguistic variations in prompts, any system relying on these predictions faces hidden instability risks.

For developers and enterprises deploying LLMs in uncertain environments, this research highlights a blind spot. The instability suggests that the same model might handle identical decision scenarios differently depending on how uncertainty is linguistically framed. This creates reproducibility and safety concerns, particularly in financial services, autonomous systems, and policy recommendation engines where consistency and transparency matter.

The research points toward urgent development needs: either refining how uncertainty is represented in LLM prompts, developing new theoretical frameworks better suited to LLM cognition, or implementing additional safeguards when deploying PT-based systems. Understanding these limitations now prevents costly failures as LLMs become more integrated into critical decision-making infrastructure.

• →Prospect Theory parameters prove unstable in LLMs when exposed to epistemic uncertainty markers, undermining theoretical reliability.
• →Current PT-based frameworks may not generalize consistently across different LLM architectures and sizes.
• →Linguistic framing of uncertainty significantly impacts LLM decision outputs, raising reproducibility concerns in production systems.
• →Deploying PT models in real-world applications with inherent ambiguity carries unquantified risk of decision inconsistency.
• →New theoretical models specifically designed for LLM decision-making may be necessary to replace PT-based approaches.