A collaborative research team has introduced a novel analytical framework that leverages large language models to decode the underlying motivations behind human decision-making. In a 2026 publication in the Proceedings of the National Academy of Sciences, the study bridges behavioral science and artificial intelligence by systematically analyzing participants free-text explanations alongside their observed choices. Led by Dr. Kamil Fuławka from the Center Synergy of Systems at TUD Dresden University of Technology, and conducted in partnership with the Max Planck Institute for Human Development and the University of Basel, the project addresses a longstanding challenge in cognitive research: translating qualitative verbal reports into scalable, rigorous quantitative data. Traditional behavioral studies rely heavily on observed actions, which often fail to capture the cognitive processes driving specific choices. To overcome this limitation, researchers designed a gambling-based experiment where participants made decisions and immediately recorded their reasoning in free text. The team developed a comprehensive taxonomy of decision rationales grounded in established economic and psychological theories, including risk aversion, loss avoidance, and outcome maximization. Large language models were then deployed to scan, classify, and extract these predefined reasons from the unstructured verbal data. Mathematical modeling of the actual choices served as an independent validation layer, ensuring that the LLM-derived insights accurately reflected participants decision patterns. The findings reveal that human decision-making is highly adaptive and context-dependent. Rather than relying on fixed cognitive strategies, individuals systematically adjust their reasoning processes based on the structural properties of the decision environment. This dynamic shift underscores the limitations of choice-only analysis and highlights the necessity of incorporating self-reported cognitive data. Dr. Fuławka emphasized the practical significance of the approach, noting that complex real-world scenarios such as financial planning, medical treatments, social dilemmas, and public policy require nuanced understanding of how people simplify information and apply heuristics. The proposed pipeline establishes a scalable methodology for behavioral research, enabling scholars to process large volumes of qualitative data with machine learning precision. By integrating natural language processing with computational modeling, the framework opens new avenues for studying human behavior in ecologically valid settings. Future applications may extend to consumer psychology, algorithmic bias detection, and personalized decision-support systems. As artificial intelligence continues to permeate cognitive science, this research demonstrates how generative models can transform subjective human narratives into actionable empirical evidence, ultimately refining our comprehension of complex decision ecosystems.