When AI Knows but Still Chooses Wrong: Associative Retrieval in LLM Financial Advice

This paper studies whether irrelevant context affects financial recommendations generated by large language models. In a controlled investment experiment, models choose between a safe asset and a risky asset after viewing payoff-irrelevant images, and separately report their probability estimates that the risky asset is good. Positive image valence raises risky choice by 16.8 percentage points, but has little effect on stated probability estimates, which remain close to Bayesian benchmarks. This belief-choice disconnect is explained by the sentiment of the model’s retrieved content and varies with their confidence and reasoning capability. Supervised fine-tuning provides further evidence for contextual retrieval: training on positive versus negative text changes risk taking, even when the training text is unrelated to finance. In a headline-based stock-news prediction task, these differences translate into different investment scores and long-short portfolio returns. The findings identify payoff-irrelevant context as a source of instability in AI-generated financial advice.