Does Code-Switching Disrupt Emotion Recognition? A Valence–Arousal Analysis of Affective Misalignment in Multilingual Human–AI Interaction

As multilingual communication increasingly involves code-switching, emotion recognition in mixed-language input has become a key challenge for human–AI interaction. This study examines emotion classification under English–Spanish and Spanish–English code-switching conditions using the valence–arousal framework. Five large language models (GPT-4 Turbo, GPT-5, Llama-3, Llama-4, and Claude Opus 4.7) were evaluated on 7,560 inputs across five code-switching ratios. Performance was assessed using accuracy, Macro-F1, and mixed-effects logistic regression, while errors were categorized as valence shifts, arousal shifts, or same-region confusions. Increasing code-switching produced a small but significant accuracy decline in both language directions. Under English-to-Spanish conditions, models generally preserved valence while showing greater arousal instability. Under Spanish-to-English conditions, GPT-5 showed increased valence-shift errors and was the only model with a significant direction–error-type association. Aggregate metrics can obscure distinct forms of affective misalignment. The proposed framework enables dimension-level comparison of error patterns across models, language directions, and code-switching ratios.