ALGORITHMIC HESITATION: EVALUATING CONFIDENCE DROPS IN LLM SENTIMENT ANALYSIS ON MIXED-EMOTION TEXT

As natural language processing (NLP) models are increasingly deployed in user-facing environments, their ability to accurately parse complex human emotions becomes critical. This paper investigates "algorithmic hesitation"—the measurable drop in classification confidence when pre-trained sentiment analysis models process mixed or ambiguous emotional text. Utilizing the Hugging Face transformers library and a pre-trained DistilBERT architecture, we subjected the model to a dataset of distinct and mixed-sentiment human statements. While the model demonstrated high baseline confidence (mean 99.88% for positive and 99.52% for negative expressions) on unilateral sentiments, confidence scores measurably decayed (to 97.77%) when processing sentences containing conflicting emotional clauses (e.g., excitement paired with dread). These findings suggest that while modern NLP pipelines are highly accurate at binary classification, their confidence intervals serve as a quantifiable metric for human cognitive dissonance, highlighting a necessary frontier for affective computing and context-aware AI.