Researchers propose testing AI systems for descriptive moral agency
An arXiv paper argues that because no settled definition of moral agency exists, researchers should instead identify and empirically test the human capabilities underlying moral judgment, starting with large language models.
In plain words
A philosophy paper asks whether AI systems can act in ways that are morally right or wrong. Since there is no agreed definition of moral agency that fits machines, the authors propose a research program: identify the human capabilities behind moral agency, refine them into testable forms, and check whether AI models, especially large language models, show them.
What was announced
The authors, writing on arXiv, note that there is no settled normative criterion of moral agency that applies neatly outside paradigm human cases, which makes it hard to judge machines by existing standards. Rather than settle that debate, they propose studying descriptive moral agency: asking what capabilities humans possess that underlie their moral agency. Drawing on historical scholarship in philosophy and moral psychology, they aim to isolate candidate capabilities and refine them so they become testable in AI systems.
The paper then offers specific empirical recommendations for testing AI systems, especially large language models, for those refined capabilities. The authors say the approach could show whether AI behaviors are aligned or broadly desirable, inform policy questions such as which models should be deployed in which contexts, complement engineering and design work by characterizing capabilities precisely, and improve understanding of the moral reasoning concepts themselves when applied to AI. The claims remain a proposal; the abstract does not report test results or name specific models evaluated.
What it does not show
- The source is only the paper's abstract, so the full methodology, experiments and findings are not visible in the supplied evidence.
- The paper proposes a research program and empirical recommendations; the supplied evidence reports no actual test results, evaluated models or measured outcomes.
- The authors' stated benefits are their own claims, not independently confirmed findings.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.