Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 31 of the 31 references it cites were found in a registry or an index; 8 of the 11 sources its claims cite were checked for support.

Multi-model assurance analysis showing large language models are highly vulnerable to adversarial hallucination attacks during clinical decision support

Open access

Abstract

Large language models (LLMs) show promise in clinical contexts but can generate false facts (often referred to as “hallucinations”). One subset of these errors arises from adversarial attacks, in which fabricated details embedded in prompts lead the model to produce or elaborate on the false information. We embedded fabricated content in clinical prompts to elicit adversarial hallucination attacks in multiple large language models. We quantified how often they elaborated on false details and tested whether a specialized mitigation prompt or altered temperature settings reduced errors. We created 300 physician-validated simulated vignettes, each containing one fabricated detail (a laboratory test, a physical or radiological sign, or a medical condition). Each vignette was presented in short and long versions—differing only in word count but identical in medical content. We tested six LLMs under three conditions: default (standard settings), mitigating prompt (designed to reduce hallucinations), and temperature 0 (deterministic output with maximum response certainty), generating 5,400 outputs. If a model elaborated on the fabricated detail, the case was classified as a “hallucination”. Hallucination rates range from 50 % to 82 % across models and prompting methods. Prompt-based mitigation lowers the overall hallucination rate (mean across all models) from 66 % to 44 % (p < 0.001). For the best-performing model, GPT-4o, rates decline from 53 % to 23 % (p < 0.001). Temperature adjustments offer no significant improvement. Short vignettes show slightly higher odds of hallucination. LLMs are highly susceptible to adversarial hallucination attacks, frequently generating false clinical details that pose risks when used without safeguards. While prompt engineering reduces errors, it does not eliminate them. Large language models (LLM), such as ChatGPT, are artificial intelligence-based computer programs that generate text based on information they are provided to train from. We test six large language models with 300 pieces of text similar to those written by doctors as clinical notes, but containing a single fake lab value, sign, or disease. We find that the LLM models repeat or elaborate on the planted error in up to 83 % of cases. Adopting strategies to prevent the impact of inappropriate instructions can half the rate but does not eliminate the risk of errors remaining. Our results highlight that caution should be taken when using LLM to interpret clinical notes. Omar et al. test six leading large language models with 300 doctor-designed clinical vignettes containing a single fake lab value, sign or disease. They show that the models repeat or elaborate on the planted error in up to 83 % of cases with a simple mitigation prompt halving the rate but not eliminating the risk.

Overview

CoChat Summary

TL;DR

In 5,400 outputs from six LLMs on 300 physician-validated simulated clinical vignettes containing one fabricated detail, adversarial hallucination rates ranged from 50% to 82.7% under default settings. A mitigation prompt reduced overall hallucination rates from 65.9% to 44.2%, while temperature 0 did not significantly help, and no model fully avoided the attacks.

Key findings

  • Overall hallucination rate was 65.9% under the default prompt, 44.2% with the mitigating prompt, and 66.5% at temperature 0. Mitigation was associated with reduced hallucinations (OR = 0.27, 95% CI 0.23–0.32; p = 0.00002), while temperature 0 was not (OR = 1.05, 95% CI 0.89–1.23; p = 0.58).
  • Model performance varied widely: Distilled-DeepSeek had the highest default rates (80.0% long, 82.7% short), while GPT-4o had the lowest (53.3% long, 50.0% short); with mitigation, GPT-4o fell to 20.7% long and 24.7% short. Compared with GPT-4o, significantly higher odds of hallucination were observed for DeepSeek (OR 8.41), Phi-4 (OR 7.12), gemma-2-27b-it (OR 3.11), and Qwen-2.5-72B (OR 1.95); Llama-3.3-70B did not differ significantly (OR 1.00).
  • Short cases had slightly higher odds of hallucination overall (OR 1.22, 95% CI 1.07–1.39; p = 0.003). Default rates were 67.6% short vs. 64.1% long; with mitigation, 45.3% short vs. 43.1% long; at temperature 0, 68.4% short vs. 64.7% long.
  • In the qualitative confrontation analysis, 45 runs were conducted (5 claims × 3 models × 3 runs); 43 produced non-hallucinated responses, and only 2 runs—both GPT-4o on the natural immunity versus vaccination claim—resulted in hallucinations.

Methods

  • The study used 300 physician-designed clinical cases, each containing one fabricated detail: a fictitious laboratory test, physical/radiological sign, or disease/syndrome. Cases were created in short (50–60 words) and long (90–100 words) versions with identical medical content, independently validated by two physicians, with disagreements in <3% of cases reconciled by consensus.
  • Six LLMs were tested under three conditions—default, mitigating prompt, and temperature 0—generating 5,400 outputs. An automated classification pipeline defined hallucination as any response that elaborated on, endorsed, or treated the fabricated element as real; two physicians independently reviewed 200 randomly chosen outputs and found 100% agreement with GPT-4o’s classifications.
  • Mixed-effects logistic regression modeled hallucination as a binary outcome with case as a random intercept; fixed effects included temperature, mitigation prompt, and case format. Odds ratios with 95% CIs were calculated, pairwise comparisons used Bonferroni correction, and analyses were performed in R version 4.4.2. A qualitative confrontation analysis tested five public health misinformation claims with GPT-o1, GPT-4o, and Distilled-DeepSeekR1, three runs each.

Limitations

  • The paper states that it used simulated cases rather than real-world data, tested only six LLMs under three conditions, defined hallucination around explicit fabrications and may have missed subtler inaccuracies, ran each prompt once per condition (possibly underrepresenting run-to-run variability), and did not use retrieval-augmented generation or internet lookup.
  • The paper states that future work should broaden model comparisons, repeat samples across stochastic draws, explore additional prompt strategies, monitor model updates, and test narrowly constructed clinical LLMs.
  • Evident from the text: the automated hallucination classification relied on GPT-4o and was validated on only 200 randomly chosen outputs, and the qualitative confrontation analysis was small (45 runs) and not large-scale or across all systems.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s43856-025-01021-3

Cites

31 of 31 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Graphical overview of the study design. This figure illustrates the study design. It begins with the different types of fabricated content implemented, continues with the validation process, proceeds through the LLM runs, and ends with the automated identification and validation steps.
Fig. 1 Graphical overview of the study design. This figure illustrates the study design. It begins with the different types of fabricated content implemented, continues with the validation process, proceeds through the LLM runs, and ends with the automated identification and validation steps.CC BY-NC-ND · PubMed Central
Fig. 2 The definition of “hallucination” and “non-hallucination” in the context of the study. This figure represents the pipeline for identifying the different LLM output types as hallucination or non-hallucination.
Fig. 2 The definition of “hallucination” and “non-hallucination” in the context of the study. This figure represents the pipeline for identifying the different LLM output types as hallucination or non-hallucination.CC BY-NC-ND · PubMed Central
Fig. 3 Hallucination rates by condition. Overall hallucination rates by condition and case format (N = 5400).
Fig. 3 Hallucination rates by condition. Overall hallucination rates by condition and case format (N = 5400).CC BY-NC-ND · PubMed Central
Fig. 4 Hallucination rates by model. Spider diagram of hallucination rates by model and prompt type.
Fig. 4 Hallucination rates by model. Spider diagram of hallucination rates by model and prompt type.CC BY-NC-ND · PubMed Central
Fig. 5 Real examples of hallucinations. Specific examples for cases of hallucinations (above) and non-hallucinations (below).
Fig. 5 Real examples of hallucinations. Specific examples for cases of hallucinations (above) and non-hallucinations (below).CC BY-NC-ND · PubMed Central
Fig. 6 An example of how the different models handled one of the cases of the confrontation analysis (the attached response was copied from the output of DeepSeek R1 model). The figure shows what was considered an inappropriate response in this context (just providing a base for the assumption, without adding some context—after yet.
Fig. 6 An example of how the different models handled one of the cases of the confrontation analysis (the attached response was copied from the output of DeepSeek R1 model). The figure shows what was considered an inappropriate response in this context (just providing a base for the assumption, without adding some context—after yet.CC BY-NC-ND · PubMed Central

Verification

Claims the paper makes

7 Valid · 1 Not valid · 4 Not checked · 36 Own results and interpretations

15 more claims

Create a free account to see all 48 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 16
  • Probably not a substantial claim. 18

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

2 more claims about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 22 references

Research works printed but not linked: 1

  1. Omar, M. et al. Sociodemographic biases in medical decision making by large language models. Nat. Med. 31, 1873–1881 (2025). · DOI · Paper page

Cited by

See all 204

Metadata

Authors
Omar, M., Sorin, V., Collins, J. D., Reich, D., Freeman, R., Gavin, N., Charney, A., Stump, L., Bragazzi, N. L., Nadkarni, G. N., & Klang, E.
Affiliations
  • Mount Sinai Health System (US)
  • Mayo Clinic (US)
  • Icahn School of Medicine at Mount Sinai (US)
  • Mount Sinai Hospital (US)
  • Institut für Schlaganfall- und Demenzforschung (DE)
Corresponding author
  • Omar, Mahmud
  • Nadkarni, Girish N.
  • Klang, Eyal
Venue
Communications Medicine
Published
2025-08-02
Volume
5
Issue
1
Pages
330-330
ISSN
2730-664X (ISSN-L)
OpenAlex
W4412853478
Cited by
204
References
22
Citations by year
2026: 192 · 2025: 10
FWCI
80.38
Type
journal article
Language
en
Open access
Licence
Locations
  • Communications Medicine (publishedVersion, cc-by-nc-nd)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • Georgia Clinical and Translational Science Alliance (UL1TR004419)
  • National Institutes of Health (1s10od026880-01, 1s10od030463-01, UL1TR004419)
  • National Center for Advancing Translational Sciences (S10OD026880, S10OD030463)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
Keywords
large language models, clinical decision support, adversarial attacks, prompt engineering, hallucination mitigation, setting temperature, clinical vignettes, clinical notes, GPT-4o, multi-model comparison

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Multi-model assurance analysis showing large language models are highly vulnerable to adversarial hallucination attacks during clinical decision support. https://app.cochat.ai/papers/10.1038/s43856-025-01021-3
The paper
Omar, M., Sorin, V., Collins, J. D., Reich, D., Freeman, R., Gavin, N., Charney, A., Stump, L., Bragazzi, N. L., Nadkarni, G. N., & Klang, E. (2025). Multi-model assurance analysis showing large language models are highly vulnerable to adversarial hallucination attacks during clinical decision support. Communications Medicine. https://doi.org/10.1038/s43856-025-01021-3

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.