<p>Artificial intelligence (AI) should be understood as political infrastructure that governs recognition, allocation, and futurity through learning-based classification and prediction. This raises the question of how AI’s political agency should be grasped when its power operates through infrastructures that prestructure what can be recognized or acted upon. Rather than treating ethics as principles applied to neutral tools, this article defines AI as learning-based governance that organizes visibility, authority, and risk. Accordingly, it frames AI ethics as political and infrastructural, understanding the political as the ethical and institutional ordering of relations among subjects and institutions. Integrating Enrique Dussel’s philosophy of liberation with Benjamin Bratton’s Stack model, it develops the Diamond Model of Political AI Ethics, which reframes AI’s ethical problems as systemic, layered, and historically embedded and grounds the analysis in Dusselian concepts of exteriority and antifetishism. The model translates four closures through which totality endures (matter, relation, being, time) into four justice dimensions (distributive, relational, ontological, temporal) and identifies valves where closure can be reopened through design, regulation, or collective action. The analysis shows how legibility, anticipation, modulation, and subjectivation convert life into computed value across AI-mediated domains. The Diamond Model contributes by locating where political agency can re-enter algorithmic systems, linking justice dimensions to Stack layers, and grounding political AI ethics in exteriority by treating those rendered illegible or uncomputable as epistemic authorities. The article thus reframes AI ethics as a politics of design aimed at keeping computational totalities incomplete.</p>