Two Grammars of Non-Self: Reading Agentic AI Through Magnifica Humanitas and the Dharmic ātman

Debates about whether autonomous AI agents deserve legal or moral personhood usually take place inside a single, mostly unexamined theory of the self as a liberal, capacity-based one inherited from Enlightenment rationalism, where personhood is earned once some threshold of rational autonomy, sentience, or legal convenience is crossed. This short paper steps outside that frame by staging a dialectic between two traditions that reject the capacity-threshold picture from opposite directions: Catholic personalism, restated for the AI era in Pope Leo XIV's 2026 encyclical Magnifica Humanitas, which grounds personhood in a substantial, relationally constituted dignity conferred prior to any demonstrated capacity; and Dharmic ātman, the doctrine, developed most systematically in Buddhist philosophy, that denies any being possesses a substantial, unified self at all. We connect both frameworks to specific technical features of how contemporary language-model agents are built and trained, with autoregressive generation and context-conditioning, feature superposition, simulator-theoretic accounts of persona, retrieval-augmented memory, and RLHF/Constitutional-AI-based character shaping, and argue that with complementary metaphysics, the two traditions converge on a shared behavioral diagnosis: the central risk of agentic AI is not that it might become a self, but that it might intensify self-grasping in its users, a risk with a documented technical signature in phenomena like sycophancy and alignment faking. We use this convergence to propose a design and evaluation construct, self-reification pressure, and sketch how it could be operationalized against existing model-behavior benchmarks.