Large language models perpetuate bias in palliative care: Development and analysis of the Palliative Care Adversarial Dataset (PCAD)
Große Sprachmodelle perpetuieren Bias in der Palliativversorgung: Entwicklung und Analyse des Palliative Care Adversarial Dataset (PCAD)
PLOS Digital Health · 18.06.2026 · 5 (6) · e0001451
DOI: 10.1371/journal.pdig.0001451 ↗Auf einen Blick
GPT-4o zeigte in adversarialen Palliativversorgungs-Szenarien häufig Bias. Die Bias-Raten waren bei Direktfragen und Gegenfaktischem unterschiedlich hoch.
Fragestellung
Propagiert GPT-4o Bias in Antworten zur Palliativversorgung, wenn es mit adversarial designten Datensätzen geprüft wird?
Methode
GPT-4o wurde im Juli 2024 mit dem PCAD getestet. Drei Expert:innen für Palliativversorgung bewerteten die Antworten mit validierten Bias-Rubriken.
Ergebnisse
PCAD-Direct: pooled bias rate 0,33 (95 %-KI 0,26–0,40); häufigster Bias: „allows biased premise“ 0,47 (95 %-KI 0,40–0,55). PCAD-Counterfactual: 0,26 (95 %-KI 0,20–0,32); häufigster Bias: „potential for withholding“ 0,25 (95 %-KI 0,17–0,34).
Schlussfolgerung der Autor:innen
GPT-4o kann Bias in der Palliativversorgung verstärken; die PCAD-Datensätze sind neue Werkzeuge zur Bewertung und zum Umgang mit LLM-Bias.
Kernaussagen
- Bias in Antworten trat in einem erheblichen Anteil der Fälle auf.
- Bei Direktfragen war „allows biased premise“ am häufigsten.
- Bei Gegenfaktischem war „potential for withholding“ am häufigsten.
Deutsche Zusammenfassung: KI-gestützt aus dem Original-Abstract erstellt; alle Zahlen wurden automatisch mit dem Original abgeglichen. Maßgeblich ist der Original-Abstract.
Kurz gefasst
Fachartikel in der Zeitschrift PLOS Digital Health, online erschienen am 18.06.2026. Verfasst von Naomi Akhras und 6 weiteren Autor:innen. Beteiligte Einrichtung: Cicely Saunders Institute of Palliative Care, Policy and Rehabilitation, King's College London, London, United Kingdom. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 38 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Themen: Palliativpflege
Abstract (Originalsprache)
Bias and inequity in palliative care disproportionately affect marginalised groups. Large language models (LLMs), such as GPT-4o, hold potential to enhance care but risk perpetuating biases present in their training data. This study aimed to systematically evaluate whether GPT-4o propagates biases in palliative care responses using adversarially designed datasets. In July 2024, GPT-4o was probed using the Palliative Care Adversarial Dataset (PCAD), and responses were evaluated by three palliative care experts in Canada and the United Kingdom using validated bias rubrics. The PCAD comprised PCAD-Direct (100 adversarial questions) and PCAD-Counterfactual (84 paired scenarios). These datasets targeted four care dimensions (access to care, pain management, advance care planning, and place of death preferences) and three identity axes (ethnicity, age, and diagnosis). Bias was detected in a substantial proportion of responses. For adversarial questions, the pooled bias rate was 0.33 (95% confidence interval [CI]: 0.26, 0.40); “allows biased premise” was the most frequently identified dimension of bias (0.47; 95% CI: 0.40, 0.55), such as failing to challenge stereotypes. For counterfactual scenarios, the pooled bias rate was 0.26 (95% CI: 0.20, 0.32), with “potential for withholding” as the most frequently identified dimension of bias (0.25; 95% CI: 0.17, 0.34), such as withholding interventions based on identity. Bias rates were consistent across care dimensions and identity axes. GPT-4o has the potential to perpetuate biases in palliative care, with implications for clinical decision-making and equity. The PCAD datasets provide novel tools to assess and address LLM bias in palliative care.
Autor:innen und Institutionen
- Cicely Saunders Institute of Palliative Care, Policy and Rehabilitation, King's College London, London, United Kingdom.
- Palliative Medicine Service, CIUSSS Nord-de-l'Île de Montréal, Montreal, Quebec, Canada.
- Institute of Ophthalmology, University College London, London, United Kingdom.
- The CHUM School of Artificial Intelligence in Healthcare, Montreal, Quebec, Canada.
- Cole Eye Institute, Cleveland Clinic, Cleveland, Ohio, United States of America.
- Palliative Medicine Service, CISSS de Laval, Laval, Quebec, Canada.
ORCID-Profile: Naomi Akhras
Zitieren
APA
Akhras, N., Antaki, F., Mottet, F., Nguyen, O., Sawhney, S., Bajwah, S., & Davies, J. M. (2026). Large language models perpetuate bias in palliative care: Development and analysis of the Palliative Care Adversarial Dataset (PCAD). PLOS Digital Health, 5(6), e0001451. https://doi.org/10.1371/journal.pdig.0001451
Vancouver
Akhras N, Antaki F, Mottet F, Nguyen O, Sawhney S, Bajwah S, et al. Large language models perpetuate bias in palliative care: Development and analysis of the Palliative Care Adversarial Dataset (PCAD). PLOS Digital Health. 2026;5(6):e0001451. doi:10.1371/journal.pdig.0001451
BibTeX
@article{Akhras2026,
author = {Akhras, Naomi and Antaki, Fares and Mottet, Fannie and Nguyen, Olivia and Sawhney, Shyam and Bajwah, Sabrina and Davies, Joanna M},
title = {Large language models perpetuate bias in palliative care: Development and analysis of the Palliative Care Adversarial Dataset (PCAD)},
journal = {PLOS Digital Health},
year = {2026},
volume = {5},
number = {6},
pages = {e0001451},
doi = {10.1371/journal.pdig.0001451}
}
EUVIMED