GPT-5’s Reasoning Process in Emotion Recognition: A Content Analysis Approach
Abstract
While Large Language Models (LLMs) demonstrate strong performance in textual emotion detection, their underlying reasoning mechanisms remain unclear. This study investigates the reasoning processes of GPT-5 by conducting a content analysis of its generated justifications for Weibo posts. Using the GPT API, we generated annotations for 3,000 posts and developed a coding scheme comprising six evidence types to manually analyze a subset of 500 justifications. The results indicate that GPT-5 employs a multi-evidence strategy, integrating surface-level linguistic cues with inferred psychological states. Correspondence analysis shows that emotions cluster into three distinct groups, each associated with specific heuristics. Furthermore, Sankey flow visualizations reveal consistent decision pathways for neutral content, contrasted with more complex and overlapping pathways for subjective emotions. Synthesizing these findings, we identify a three-stage reasoning mechanism: (1) factual filtering, where the model separates objective information from emotional expression; (2) stance projection, prioritizing evaluative stance and tone over explicit emotion words to infer emotions such as anger; and (3) contextual simulation, linking situational events with inferred psychological states to interpret emotions such as sadness or fear. This study provides an empirical framework for interpreting LLM emotional reasoning, shifting the focus from performance evaluation to the interpretability of reasoning processes.