Implicit bias in digital health: systematic biases in large language models' representation of global public health attitudes and challenges to health equity.

Introduction: As emerging instruments in digital health, large language models (LLMs) assimilate values and attitudes from human-generated data, thereby possessing the latent capacity to reflect public health perspectives. This study investigates into the representational biases of LLMs through the lens of health equity. We propose and empirically validate a three-dimensional explanatory framework encompassing Data Resources, Opinion Distribution, and Prompt Language, positing that prompts are not just communicative media but critical conduits that embed cultural context. Methods: Utilizing a selection of prominent LLMs from the United States and China-namely Gemini 2.5 Pro, GPT-5, DeepSeek-V3, and Qwen 3. We conduct a systematic empirical analysis of their performance in representing health attitudes across diverse nations and demographic strata. Results: Our findings demonstrate that: first, the accessibility of data resources is a primary determinant of an LLM's representational fidelity for internet users and nations with high internet penetration. Second, a greater consensus in public health opinion correlates with an increased propensity for the models to replicate the dominant viewpoint. Third, a significant "native language association" is observed, wherein Gemini 2.5 Pro and DeepSeek-V3 exhibit superior performance when prompted in their respective native languages. Conversely, models with enhanced multilingual proficiencies, such as GPT-5.0 and Qwen 3, display greater cross-lingual consistency. Discussion: This paper not only quantifies the degree to which these leading LLMs reflect public health attitudes but also furnishes a robust analytical pathway for dissecting the underlying mechanisms of their representational biases. These findings bear profound implications for the advancement of health equity in the artificial intelligence era.