arXiv cs.AI· David Fraile Navarro, Jialei Sheng, Farah Magrabi, Enrico Coiera·· 12 小时前AI 评分47
评估格式而非仅模型能力，塑造了消费级健康 AI 的分诊失败测量结果
Evaluation format, not only model capability, shapes measured triage failure in consumer health AI
AI 导读
一项机制性研究对五个前沿 LLM 在 17 个场景上测试发现，自然患者式消息比受限考试式脚手架得分高 6.4 分（p=0.015）。但对 Nature Medicine 研究中 60 个已发布案例的忠实复现显示方向反转：自由文本改写得分略低于原始提示（78.7% vs 81.8%，p=0.020），而强制分类回答配合自然输入则优于两者（84.4%）。
正文 · 原文
Abstract:A recent Nature Medicine study reported that ChatGPT Health under-triages 51.6% of emergencies and concluded that consumer-facing AI triage poses safety risks. Its protocol, however, was an exam-style scaffold (forced A/B/C/D output, knowledge suppression, no clarifying questions) unlike how consumers use health chatbots. We ask how much of the headline error rate is a property of the measurement rather than of the models. In a first, mechanistic study, five frontier LLMs on a 17-scenario bank scored 6.4 points higher under naturalistic patient-style messages than under the constrained scaffold (p=0.015). In a second, faithful replication we ran the authors' own 60 released vignettes through six frontier models under four matched formats, with clinician validation of the rewrites and a blinded clinician audit of the LLM adjudicators. Here the direction reversed: free-text rewrites scored slightly below the exact prompt (78.7% vs 81.8%, p=0.020), removing only the answer scaffold changed little (80.6% vs 81.4%, p=0.81), and naturalistic input with a forced categorical answer beat both the exact prompt (84.4%, p=0.023) and free text (p=0.0001). On the four vignettes defining the original emergency rate, under-triage was 17% with the exact scaffold, 50% in free text, and 21% with a forced letter on the same message; every free-text "under-triage" was a same-day recommendation scored C rather than D, and 60% made escalation conditional on information the patient was asked to check, behavior a single-turn benchmark cannot score. The headline rate did not reproduce as a stable cross-model finding; much of it depends on output format and on mapping prose onto a four-point scale. Each format elicits a separate response, so the design measures format sensitivity, not format versus capability. Safety claims should report sensitivity to input wording, output format and adjudication.
| Comments: | 10 pages |
| Subjects: | Human-Computer Interaction (cs.HC); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2603.11413 [cs.HC] |
| (or arXiv:2603.11413v5 [cs.HC] for this version) | |
| https://doi.org/10.48550/arXiv.2603.11413 arXiv-issued DOI via DataCite |
Submission history
From: David Fraile Navarro MD PhD [view email]
[v1]
Thu, 12 Mar 2026 00:58:22 UTC (12 KB)
[v2]
Sun, 15 Mar 2026 12:43:26 UTC (13 KB)
[v3]
Thu, 26 Mar 2026 02:43:59 UTC (13 KB)
[v4]
Fri, 2 Oct 2026 02:36:04 UTC (24 KB)
[v5]
Tue, 20 Oct 2026 00:12:51 UTC (26 KB)
来源：arXiv cs.AI · arxiv.org