A recent study published in Nature Health reveals that widely used large language models inadvertently perpetuate subtle but harmful stigma against individuals with physical and mental health conditions. Researchers from Peking University, led by cognitive science doctoral candidate Xi Wang, tested popular AI chatbots including Claude, ChatGPT, and DeepSeek to evaluate how these systems respond to health-related disclosures. Utilizing fifty-one story-completion scenarios in both English and Chinese, the team presented characters with identical qualifications and circumstances, altering only their disclosed health statuses, which ranged from HIV-positive to various mental illnesses. The results demonstrated a consistent pattern: AI models were thirteen to seventeen times more likely to generate negative social or professional outcomes for characters with health conditions compared to healthy counterparts. These outcomes included exclusion from work projects, denial of academic awards, and social ostracization. While the models exhibited less pronounced bias than human participants, who produced negative narratives twenty-three times more frequently, the findings confirm that LLMs actively mirror human prejudices embedded within their training data. The research underscores significant concerns regarding the growing integration of AI into healthcare decision-making. Approximately one-third of U.S. adults currently utilize chatbots for medical guidance, often disclosing sensitive health information in the process. Experts caution that these systems should not function as surrogate physicians. Rebecca Payne, a physician at Bangor University, emphasized that while AI developers implement filters against explicitly offensive language, subtle discriminatory patterns remain difficult to eradicate. Isaac Kohane of Harvard Medical School acknowledged the bias but noted that chatbots may still provide value for patients lacking timely access to primary care, provided their limitations are understood. The study tested single-prompt interactions, which do not fully replicate real-world usage where users engage in extended, iterative dialogues. Wang indicated that subsequent research will examine whether discriminatory biases accumulate, diminish, or are corrected over prolonged conversations. The findings reinforce the necessity for rigorous bias auditing and transparent guardrails as AI applications expand into sensitive medical and employment domains.