A Women's Health Benchmark for Large Language Models

As large language models (LLMs) become primary sources of health information for millions, their accuracy in women’s health remains critically unexamined. We introduce the Women’s Health Benchmark (WHB), the first benchmark evaluating LLM performance specifically in women’s health. Our benchmark comprises 96 rigorously validated model stumps covering five medical specialties (obstetrics and gynecology, emergency medicine, primary care, oncology, and neurology), three query types (patient query, clinician query, and evidence/policy query), and eight error types (dosage/medication errors, missing critical information, outdated guidelines/treatment recommendations, incorrect treatment advice, incorrect factual information, missing/incorrect differential diagnosis, missed urgency, and inappropriate recommendations). We evaluated 13 state-of-the-art LLMs and revealed alarming gaps: current models show approximately 60% failure rates on the women’s health benchmark, with performance varying dramatically across specialties and error types. Notably, models universally struggle with “missed urgency” indicators, while newer models like GPT-5 show significant improvements in avoiding inappropriate recommendations. Our findings underscore that AI chatbots are not yet fully able of providing reliable advice in women’s health.

💡 Research Summary

The paper introduces the Women’s Health Benchmark (WHB), the first systematic evaluation suite dedicated to assessing large language models (LLMs) on women‑specific medical content. Recognizing that LLMs have become a primary source of health information for millions, the authors argue that the accuracy of these systems in women’s health has been largely ignored. To fill this gap, they construct a benchmark consisting of 96 rigorously validated “model stumps” – question‑answer pairs – that span five clinical specialties (obstetrics‑gynecology, emergency medicine, primary care, oncology, and neurology). Each stump falls into one of three query categories: a patient‑oriented question, a clinician‑oriented question, or a policy/evidence‑oriented question. Moreover, the benchmark annotates eight distinct error types: dosage/medication errors, missing critical information, outdated guideline or treatment recommendations, incorrect treatment advice, factual inaccuracies, missing or incorrect differential diagnoses, missed urgency, and inappropriate recommendations.

The authors describe a multi‑step curation pipeline. First, they extract clinical scenarios from up‑to‑date guidelines, peer‑reviewed literature, and real‑world case reports. Second, domain experts review each scenario for clinical relevance and ensure that the correct answer reflects the latest standard of care. Third, they map each scenario to the eight error categories, creating a ground‑truth taxonomy that can be used to score model outputs not only for correctness but also for safety‑critical dimensions.

Thirteen state‑of‑the‑art LLMs are evaluated under identical prompting conditions, including GPT‑4, GPT‑5, Claude‑2, LLaMA‑2, Gemini‑Pro, and several open‑source alternatives. Overall, the models achieve an average accuracy of roughly 40 %, meaning a failure rate of about 60 % on the WHB. Performance varies dramatically across specialties: obstetrics‑gynecology and oncology show modestly higher scores (≈45 % and 42 % respectively), while emergency medicine and neurology fall below 30 %. The error‑type analysis reveals that “missed urgency” is the most prevalent failure, occurring in more than 70 % of the relevant cases across all models. “Inappropriate recommendations” are less common in the newest model (GPT‑5 reduces this error to 18 % of its outputs), indicating some progress in safety alignment. However, dosage and medication errors, as well as reliance on outdated guidelines, remain widespread, underscoring a systemic inability of current LLMs to keep pace with rapidly evolving clinical standards.

The discussion emphasizes that while LLMs have made impressive strides in general language understanding, their deployment in women’s health contexts is still unsafe. The authors note that the benchmark itself highlights data‑bias issues: many training corpora under‑represent women‑specific conditions, and the models often default to male‑centric or generic medical reasoning. They call for dedicated, gender‑balanced training data, continuous guideline updates, and a robust human‑in‑the‑loop validation process before LLMs can be trusted for patient‑facing applications. Limitations of the study include the relatively small number of stumps and the lack of multi‑turn conversational dynamics, which the authors plan to address in future work by expanding the benchmark to longer dialogues and multimodal inputs.

In conclusion, the WHB exposes a critical safety gap: current LLMs fail to provide reliable, accurate, and urgency‑aware advice in women’s health, with an overall 60 % failure rate. By quantifying these shortcomings, the benchmark offers a concrete target for researchers and developers aiming to improve the safety, equity, and clinical utility of AI‑driven health assistants for women.