Introduction. Large language models (LLMs) are increasingly applied in Chinese sentiment analysis; however, their ability to interpret sentiment in culturally specific contexts remains uncertain. This study systematically evaluates the potential biases of LLM-based Chinese sentiment analysis, providing an empirical basis for model optimisation and responsible application. Method. A human-AI comparison experiment was conducted at both word and sentence levels. Two hundred Chinese words from four categories and 225 sentences were used as materials. Judgment data on sentiment polarity, intensity, valence, and arousal were collected from native Chinese speakers and representative Chinese and international LLMs. Data analysis was performed using Chi-square tests and Kruskal–Wallis tests. Results. In sentiment polarity judgment, LLMs are highly consistent with humans and outperform the traditional sentiment lexicon. For continuous dimensions (intensity, valence, arousal), LLMs generally show an exaggeration bias, especially for sensory-perceptual words and ironic sentences. The bias stems from the disembodied cognitive nature of models, exaggerated linguistic patterns in training data, and technical characteristics such as the attention mechanism. Conclusion. This study reveals a systematic limitation of LLMs in Chinese sentiment analysis, characterised by accurate classification but exaggerated quantitative evaluations. Consequently, in practical applications, outputs related to sentiment intensity should be interpreted with caution.