Uncovering the reasons behind our decisions
Large language models can evaluate people’s written explanations of their choices at scale, accurately identifying the reasons behind their decisions
Researchers at the Max Planck Institute for Human Development have investigated how large language models can be used to identify the reasons people give for their choices. The models make it possible to systematically evaluate large volumes of open-ended responses. This new approach could improve the understanding of decision-making processes and pave the way for more context-sensitive models of human decision making.
Muesli or toast for breakfast? Stick with a steady job or take a chance on an exciting offer? Hop on the bike or grab the car keys? We are constantly making decisions—some small, some large; some on autopilot, others only after a long process of deliberation. The outcome of a decision is plain to see. Getting to the bottom of why someone chose it—and which considerations tipped the balance—is another matter entirely.
Researchers at the Center for Adaptive Rationality of the Max Planck Institute for Human Development in Berlin have been looking into these questions. In a recent study, they examined the reasons people gave for their decisions—and whether large language models were able to correctly identify those reasons in decision-makers’ own verbal reports.
The study was motivated by a long-standing methodological challenge: People’s verbal reports on how and why they made a decision have long been considered an unreliable source of data in psychology and economics. Decision research has instead relied on “hard” behavioral data: the actual choices people make. Yet choice alone does not cast light on the reasons behind people’s decisions. And without knowledge of those reasons, the decision-making processes underlying a choice—and how they vary across people and situations—remain hidden.
Against this background, the research team, led by psychologist Dirk Wulff of the Center for Adaptive Rationality, took another approach. They presented study participants with a choice task and then asked them to describe in their own words how and why they made their decision. When datasets are large, evaluating open-ended responses of this kind takes considerable time and effort. The question is whether large language models make it possible to systematically analyze such responses at scale.
How accurately does AI identify reasons for decisions?
The researchers first had to test whether the models were capable of accurately identifying participants’ stated reasons. To this end, they developed a set of 47 possible reasons, describing each both verbally and in the form of an unambiguous decision rule. They then tested whether eight large language models were able to correctly apply a stated reason to a specific decision. The best-performing model accurately identified the chosen option in over 96% of cases. The researchers subsequently used this model to analyze the participants’ reports.
The researchers had 86 adults complete 20 monetary choice problems. In each case, they chose between two monetary lotteries, each offering different outcomes with different probabilities. For example, participants might be asked to choose between one lottery offering a 0.1% chance of winning €6,000 and another lottery offering a 0.2% chance of winning €3,000. Both lotteries paid out nothing if they didn’t win. After each choice, participants were asked to describe in their own words how and why they made their decision. This yielded 1,720 short verbal reports on the reasons for specific decisions.
The large language model was then instructed to examine the verbal reports to determine which of the previously defined decision reasons could be identified in them. For each reason identified, it rated on a scale from 0 to 100 how confident it was that the reason appeared in the respective report.
The researchers were then able to examine whether the decision reasons identified by the large language model predicted participants’ choices. Crucially, the identified decision reasons aligned with participants’ actual choices in 95% of trials. This does not mean that the model reconstructed participants’ thoughts. However, the identified reasons corresponded to a high degree with their actual choices.
Decision reasons vary more strongly with context than across individuals
Further analysis showed that the decision reasons used were by no means the same across tasks. Which reasons played a role differed more between choice problems than between individual participants. In other words, participants did not fall back on the same fixed strategy across decisions. Instead, their reasons depended strongly on the structure of the problem—that is, on the constellation of gains, losses, and probabilities at stake.
The researchers then examined whether the identified decision reasons predicted people’s actual choices in further decisions. To this end, they developed predictive models based on the previously found patterns. The models’ predictions proved to be competitive with those of established computational models in decision research. The authors conclude that verbal reports contain information that cannot be obtained from observed choices alone.
By pinpointing the reasons behind people’s decisions in different situations, researchers can bring to light previously hidden differences in decision-making processes. Such information can help test existing theories, expose their weaknesses, and guide the development of more context-sensitive models. The authors therefore see the approach not only as a new method of analysis, but also as a tool for scientific discovery—one that could, in principle, also be applied in other domains, such as intertemporal, moral, or social decision making.
Future decision research could then not only assess choices more precisely, but also explain how different situations give rise to different decision-making processes. This would make it possible for human decision behavior to be explained more fully—and predicted more reliably.
Original publication
Fuławka, K., Hertwig, R., & Wulff, D. U. (2026). Large language models accurately identify decision reasons in verbal reports. Proceedings of the National Academy of Sciences of the United States of America, 123(27), Article e2526798123. https://doi.org/10.1073/pnas.2526798123












