The Ethics in the Mirror: A Genealogical Test for Moral Agency Attribution in Algorithmic Systems

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Artificial intelligence systems increasingly produce actions that humans classify as morally salient. They refuse or comply with harmful instructions, use tools, navigate digital environments, select multi-step trajectories, and increasingly produce consequential actions in physical environments. Such observations are essential for safety evaluation. They do not, however, independently establish the moral agency of the systems producing them. This paper introduces the Moral Agency Attribution Gap (MAAG): the epistemic interval between the normative classification of observed algorithmic behavior and the justified attribution of moral agency to the system producing it. The contribution is not the distinction between behavior and moral agency itself, which is already present in the literature, but the specification of an intermediate epistemic procedure for determining whether observations of the former warrant attribution of the latter. We propose the Genealogical Test for Moral Agency Attribution (GTMA). The test requires: (1) observation and normative classification of a target behavior; (2) reconstruction of its available causal-normative genealogy across training, objectives, reinforcement structures, safety constraints, instructions, permissions, environmental provenance, embodiment, and operational jurisdiction; (3) assessment of the explanatory sufficiency of that genealogy through genealogical coverage, mechanistic plausibility, and counterfactual sensitivity; and (4) investigation of any structured Normative Residual that remains insufficiently explained. The framework neither assumes nor denies the possibility of artificial moral agency. Human provenance does not imply complete human determination: algorithmic transformation and generalization may generate trajectories that no individual human actor explicitly specified or anticipated. GTMA instead establishes an epistemic ordering. Moral agency should not be introduced as an explanatory hypothesis for a particular observation while that observation remains adequately explained by reconstructed provenance, operational conditions, and algorithmic transformation. The resulting framework distinguishes material danger, operational agency, normative classification, moral responsibility, and moral agency without reducing one to another. It reframes the question from whether an artificial action appears moral or immoral to what evidence would justify attributing the normative origin of that action to the artificial system itself.