The Limits of AI in Understanding Emotions: Challenges in Bridging Human Experience and Machine Perception

Background This article explores the potential of Artificial Intelligence (AI) to replace the human factor in the analysis of emotions within financial decision-making contexts. The background of this study is as follows: growing reliance on Artificial Intelligence for text analysis, understanding its capacity to detect emotional patterns is particularly relevant in fields where emotional dynamics significantly influence behavior, such as trading. Method Our research method is based on a three-day trading experiment involving students, during which participants made decisions under conditions of uncertainty. At the end of the experiment, each participant wrote a free-form report describing their emotional experience. We applied four different Artificial Intelligence tools to analyze these texts and compared the results with those obtained through a traditional lexical analysis method. AI tools provided relatively consistent and coherent assessments of the general emotional tone in the texts. Results Our results suggest that they successfully identified dominant emotions such as fear, disappointment, and hope. However, the ability of Artificial Intelligence to detect more nuanced or mixed emotional states was limited. In contrast, traditional lexical analysis offered greater sensitivity to emotional complexity. Although both approaches converged on broad emotional tendencies, divergences appeared when addressing subtle or context-specific emotional shifts. Our findings suggest that AI can serve as a useful preliminary tool in the analysis of emotions in written texts, especially for identifying dominant patterns. Conclusion In conclusion, according to us, due to its limitations in detecting emotional nuances, Artificial Intelligence should be integrated into a hybrid analytical approach. Combining Artificial Intelligence with traditional or human-led analysis enhances both the precision and depth of interpretation. We argue that a hybrid model avoids the oversimplification of emotional data and better reflects the complexity of emotional dynamics in decision-making under uncertainty.