Introduction
Why do people make the choices they do? This fundamental question has long intrigued researchers across psychology, economics, and neuroscience. Traditional approaches often rely on observing choices in controlled experiments and inferring underlying decision processes through mathematical models. However, these methods can only capture what people do, not why they do it. Now, a team of researchers from the Center Synergy of Systems (SynoSys) at TUD Dresden University of Technology, the Max Planck Institute for Human Development, and the University of Basel has developed a novel approach that combines observed choices with participants' own descriptions of their decision processes. By leveraging large language models (LLMs), they can systematically analyze free-text explanations to uncover the reasons behind human decisions in unprecedented detail. Their findings are published in the Proceedings of the National Academy of Sciences.
Turning Explanations into Data
Lead author Dr. Kamil Fuławka, a researcher at SynoSys, explains: "Our understanding of human behavior, including decision-making, can be deepened by asking people to elaborate on their thought processes. However, the systematic analysis of such free-text data requires scalable and rigorous analytical frameworks—an endeavor that can now be supported by LLMs." In their experiment, participants engaged in a gambling task and were asked to explain each decision in their own words. To analyze these explanations, the researchers drew on existing theories and models of decision-making to develop a large set of possible decision reasons, such as focusing on the best possible outcome or avoiding a big loss. Large language models then identified which of these reasons appeared in participants' free-text explanations, while mathematical modeling of people's choices provided validation.
How the Study Worked
The study involved a controlled gambling experiment where participants made a series of decisions between risky options. After each choice, they typed a brief explanation of their reasoning. The researchers compiled a comprehensive list of potential decision reasons based on established decision-making theories, including:
- Maximizing expected value
- Minimizing potential losses
- Seeking the highest possible payoff
- Avoiding the worst outcome
- Following a simple heuristic like "choose the option with the higher probability of winning"
Using a large language model, they automatically classified each free-text response to determine which reasons were present. To ensure accuracy, the LLM's classifications were cross-validated with human raters and compared against mathematical models that predicted choices based on the identified reasons. This multi-method approach allowed the team to both discover the reasons people cite and verify that those reasons actually drive their decisions.
Decision Reasons Shift with Context
The combination of verbal reports, LLMs, and rigorous mathematical modeling clearly demonstrated that people's own insights are a valuable source of data. It also showed that the reasons people rely on are not fixed but shift systematically with the structure of the decision problem. For example, when the potential losses were large, participants were more likely to mention loss avoidance; when the potential gains were high, they focused on maximizing gains. This context-dependence challenges traditional models that assume stable preferences or fixed decision strategies.
"Many important decisions—from financial investments to medical choices—involve weighing risks and benefits," says Dr. Fuławka. "Our method reveals that people adapt their reasoning to the specific situation, which has implications for predicting behavior and designing interventions."
Implications for Behavioral Science
This research opens new avenues for studying human behavior. By integrating free-text data with LLMs, scientists can now access the rich, qualitative information that participants naturally provide, without the constraints of predefined survey questions. The approach is scalable, allowing analysis of thousands of responses quickly and consistently. Moreover, it provides a bridge between qualitative and quantitative methods, offering a more complete picture of decision-making.
The study also highlights the potential of LLMs as tools for behavioral research. While LLMs are often criticized for lacking true understanding, here they serve as powerful pattern-matching engines that can reliably detect decision reasons in text. The researchers emphasize that the LLM's classifications were validated against human judgments and mathematical models, ensuring reliability.
Future Directions
The team plans to apply their method to other domains, such as consumer choice, political decision-making, and health behavior. They also aim to refine the set of decision reasons and explore how individual differences (e.g., age, cognitive ability) influence the reasons people use. Ultimately, this work could lead to more accurate models of human choice that incorporate both behavioral data and self-reported reasoning.
As LLMs continue to improve, their ability to analyze complex human language will only grow. This study demonstrates a practical application that respects the richness of human introspection while maintaining scientific rigor. The hidden reasons behind our choices may no longer be so hidden.
This article is based on reporting by Phys.org. Read the original article.
Originally published on phys.org
