A recent study published in Nature Health reveals a critical disparity in how large language models evaluate health-related stigma, demonstrating that AI systems often perform well on standard fairness metrics while still generating discriminatory outputs in real-world scenarios. Researchers tested six leading models, including ChatGPT, Grok, and Claude, using a dual-method approach to assess bias across conditions such as HIV, hepatitis B, and mental illness. When administered conventional stigma questionnaires, the models scored significantly lower in prejudice than human benchmarks. However, when tasked with completing 51 contextual story scenarios where only the subject’s medical condition varied, the models exhibited pronounced bias. The contextual evaluation exposed how AI silently reshaped narratives based on diagnosis. Models frequently associated HIV and mental health conditions with danger and social distancing, while attributing chronic physical ailments like hypertension to pity or perceived incompetence. Language played a notable role, with Chinese prompts yielding higher stigma-congruent responses than English, particularly in mental health contexts. Notably, enforcing step-by-step reasoning before responses substantially reduced biased outputs. These findings underscore a persistent vulnerability in medical AI deployment. As chatbots increasingly assist with patient screening, care coordination, and therapeutic support, implicit biases can directly influence diagnostic trust and care-seeking behavior. Traditional fairness assessments, which rely on direct questioning, fail to capture how models contextualize information during dynamic interactions. The study demonstrates that explicit compliance with bias surveys does not guarantee equitable performance in practical applications. To bridge this gap, researchers outlined nine mitigation strategies, emphasizing prompt engineering techniques such as individualization and relevance filtering, which instruct models to disregard irrelevant health status. The authors recommend that healthcare institutions adopt standardized prompting toolkits and that AI developers integrate contextual judgment audits into prerelease evaluation pipelines. Implementing these measures is essential to prevent algorithmic discrimination, preserve patient trust, and ensure that generative AI supports rather than compromises equitable healthcare delivery.