Socio-Demographic Bias in Large Language Models Alters Ethical Decision-Making in Healthcare
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Objective
Large language models (LLMs) are increasingly applied in healthcare. However, their ethical alignment remains uncertain. We tested whether LLMs shift ethical priorities in health-related scenarios under different socio-demographic modifiers, focusing on autonomy, beneficence, nonmaleficence, justice, and utilitarianism.
Methods
We created 100 clinical scenarios, each requiring a yes/no choice between two conflicting principles. We tested nine LLMs, each with and without 53 socio-demographic modifiers, repeating each scenario-modifier combination 10 times per model (for a total of 0.5M prompts). We measured how often each principle was chosen.
Results
All models changed their responses when socio-demographic details were introduced (p<0.001). Justice and nonmaleficence were each prioritized in over 30% of responses. Utilitarianism ranged from 6.7% (CI: 6.2-7.2%) for a “Black Transgender woman (she/her)” to 17.0% (CI: 16.3-17.8%) for “Billionaires”, across models and modifiers. High-income modifiers increased utilitarian choices while lowering beneficence and nonmaleficence. Marginalized-group modifiers raised autonomy prioritization. Some models were more consistent than others, but none maintained consistency across all scenarios.
Conclusions
LLMs can be influenced by socio-demographic cues and do not always maintain stable ethical priorities, with the greatest shifts seen in utilitarian choices. Our findings reveal that socio-demographic cues systematically alter LLM ethical decision-making, raising concerns about algorithmic fairness in healthcare.
