Can LLM-Based Digital Twins Approximate Human Loss Aversion and Reward Anticipation in Consumer Decision-Making? A Narrative Review of Neural Benchmarks and Computational Limitations

Abstract

Background: Consumer neuroscience has identified loss aversion and reward anticipation as two core mechanisms shaping purchase decisions, but traditional measurement methods such as fMRI and EEG remain expensive and difficult to scale. Methods: This narrative review synthesizes peer-reviewed literature on prospect theory, dopaminergic reward circuitry, consumer neuroscience, and recent work on large language model (LLM) based digital twins and silicon sampling to evaluate whether LLMs can approximate these mechanisms. Results: One recent consumer digital-twin study reported 86% future-purchase prediction accuracy using individualized fine-tuning on Amazon transaction data, and a large-scale behavioral benchmark found that even the best-performing LLMs achieved only modest simulation fidelity (40.80/100) across 20 diverse human-behavior datasets. Available studies have not established that LLMs replicate the behavioral asymmetry of loss aversion, the temporal dynamics of reward anticipation, or individual variation in these mechanisms. Conclusion: Current evidence suggests that LLMs function as useful behavioral prediction tools for consumer research but have not yet been shown to reproduce the mechanistic processes documented in consumer neuroscience.