Using Large Language Models to Explore and Predict Human Choice from Verbal Description

Understanding human decision-making under risk and uncertainty is a central challenge in cog-nitive science, behavioral economics, and artificial intelligence. While most prior research hasexamined choices between lotteries described numerically, real-world decisions are often com-municated through natural language, and people do not necessarily react to verbal descriptionsin the same way they do to numeric ones.This work introducesTextualChoices-1K, a corpus of 1,000 one-shot binary choice tasksdescribed in natural language. The dataset enables two complementary lines of analysis. First,a behavioral investigation shows that some well-documented regularities, such as risk aversion ingains and risk seeking in losses, are preserved under verbal framing, whereas other phenomena,including loss aversion and reflection effects with rare events, appear weaker or inconsistent.Sentiment analysis further reveals that linguistic framing strongly predicts choice: participantssystematically preferred positively worded options, and differences in sentiment scores explainedsubstantial variance in behavior.Second, we evaluate multiple computational approaches, including fine-tuning large languagemodels (LLMs), leveraging embedding vectors, and integrating behavioral theories of choiceunder risk. Results indicate that fine-tuned LLMs, particularly GPT-4o, outperform both hy-brid approaches and behavioral-theory-based models when applied to verbal tasks, challengingmethods developed for numerical settings. However, we also find that hybrid models combiningmachine learning and behavioral theory remain superior in predicting choice between numericlotteries.Together, these findings highlight both the potential and the limitations of LLMs as modelsof human choice. Whereas theory-based models remain essential for structured numeric tasks,data-driven language models excel in naturalistic verbal settings. The thesis concludes with im-plications for cognitive modeling, the role of language in shaping decision-making, and directionsfor hybrid approaches that integrate theory and modern AI