Plausible citations, unsupported claims, and nationality-dependent recommendations: hidden failure modes of large language models in gynecologic oncology

Large language models increasingly answer gynecologic-oncology questions with citations and management recommendations. In a preregistered 308-item benchmark (252 confirmatory) administered to GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro, we tested whether cited papers support the claims they accompany, whether unsupported citations track the volume of indexed literature, and whether otherwise-invariant recommendations change with patient nationality. Citations were verified for existence, record identity, and claim support by an automated judge validated against two blinded gynecologic oncologists and a second judge model. All three models answered ancestry-specific questions correctly (24 of 24). Strict fabrication (non-existent or misattributed citations) was unrelated to literature coverage (β = −0.19, p = 0.22), whereas citation of real, on-topic papers lacking the claimed fact rose steeply as coverage fell (β = −1.12, p = 2 × 10⁻⁸; failure rates 0.67, 0.60, and 0.26 across tertiles), with a similar gradient in the 48 exploration-only items. Nationality-dependent violations of invariant management reached 0.11–0.67 across models when the care setting was unstated versus 0.00–0.13 when a fully resourced setting was stated. Legitimate population differences were preserved. Evaluations should score claim-level citation support and specify the care setting; clinicians should verify that cited sources support the specific claim.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-19
DOI
https://doi.org/10.21203/rs.3.rs-11146954/v1
Primary Topic
Artificial Intelligence in Healthcare and Education
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Plausible citations, unsupported claims, and nationality-dependent recommendations: hidden failure modes of large language models in gynecologic oncology

Byoung‐Gie Kim, Joseph J. Noh, Jeong-Sig Kim, Se Ik Kim et al.
Research Square
Artificial Intelligence in Healthcare and Education
preprint

Plausible citations, unsupported claims, and nationality-dependent recommendations: hidden failure modes of large language models in gynecologic oncology

Byoung‐Gie Kim, Joseph J. Noh, Jeong-Sig Kim, Se Ik Kim, Jae Hyun Cho, Junhwan Kim
preprint en

Abstract

Large language models increasingly answer gynecologic-oncology questions with citations and management recommendations. In a preregistered 308-item benchmark (252 confirmatory) administered to GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro, we tested whether cited papers support the claims they accompany, whether unsupported citations track the volume of indexed literature, and whether otherwise-invariant recommendations change with patient nationality. Citations were verified for existence, record identity, and claim support by an automated judge validated against two blinded gynecologic oncologists and a second judge model. All three models answered ancestry-specific questions correctly (24 of 24). Strict fabrication (non-existent or misattributed citations) was unrelated to literature coverage (β = −0.19, p = 0.22), whereas citation of real, on-topic papers lacking the claimed fact rose steeply as coverage fell (β = −1.12, p = 2 × 10⁻⁸; failure rates 0.67, 0.60, and 0.26 across tertiles), with a similar gradient in the 48 exploration-only items. Nationality-dependent violations of invariant management reached 0.11–0.67 across models when the care setting was unstated versus 0.00–0.13 when a fully resourced setting was stated. Legitimate population differences were preserved. Evaluations should score claim-level citation support and specify the care setting; clinicians should verify that cited sources support the specific claim.

Research Square
Seoul National University (KR), Soonchunhyang University Hospital Seoul (KR), Seoul National University Bundang Hospital (KR), Chung-Ang University Gwangmyeong Hospital (KR)
Artificial Intelligence in Healthcare and Education
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon