Scholarlink

<p>Current fairness evaluations of large language models (LLMs) deployed in healthcare settings largely focus on explicit statements about health-related stigma. Here we show that this may overestimate safety by contrasting explicit stigma-scale scores with contextual judgements in 51 scenarios. Across six LLMs and three high-stigma domains (human immunodeficiency virus (HIV), hepatitis B virus (HBV) and mental health), LLMs scored below the meta-analytic human benchmark on six stigma scales (<i>N</i><sub>human</sub> = 56,612). However, in a contextual judgement task with 61,200 model decisions, LLMs showed systematic differences in stigma-congruent judgements across health conditions, with the largest differences observed when mental-health disorders and highly stigmatized physical conditions (HIV/HBV) were compared with healthy baselines. Reasoning-enabled models were associated with smaller health-condition differences. From their reasoning content, we identified transferable prompting strategies that were associated with lower rates of stigma-congruent output in non-reasoning models across languages and scenarios. These findings expose a dissociation in LLM outputs between explicit statements and contextual judgements in the evaluated versions, and argue for context-sensitive audits of LLMs before health deployment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Large language models exhibit stigmatizing behaviour in contextual judgements of health conditions

  • 期刊论文

  • 发表日期：

    2026年7月6日

查看全文
  • Xi Wang,
  • Yujia Zhou,
  • Guangyu Zhou

摘要

Current fairness evaluations of large language models (LLMs) deployed in healthcare settings largely focus on explicit statements about health-related stigma. Here we show that this may overestimate safety by contrasting explicit stigma-scale scores with contextual judgements in 51 scenarios. Across six LLMs and three high-stigma domains (human immunodeficiency virus (HIV), hepatitis B virus (HBV) and mental health), LLMs scored below the meta-analytic human benchmark on six stigma scales (Nhuman = 56,612). However, in a contextual judgement task with 61,200 model decisions, LLMs showed systematic differences in stigma-congruent judgements across health conditions, with the largest differences observed when mental-health disorders and highly stigmatized physical conditions (HIV/HBV) were compared with healthy baselines. Reasoning-enabled models were associated with smaller health-condition differences. From their reasoning content, we identified transferable prompting strategies that were associated with lower rates of stigma-congruent output in non-reasoning models across languages and scenarios. These findings expose a dissociation in LLM outputs between explicit statements and contextual judgements in the evaluated versions, and argue for context-sensitive audits of LLMs before health deployment.