Counting Another's Treasure: Can a Large Language Model Understand the Dharma?

Abstract

The debate over whether large language models (LLMs) "understand" has stalled because "understanding" is treated as a single, intrinsic property that a system either has or lacks. This paper imports a finer instrument from Buddhist epistemology: the threefold cultivation of wisdom as śrutamayī (wisdom from hearing), cintāmayī (wisdom from reflection), and bhāvanāmayī (wisdom from cultivation), together with the two means of valid cognition (pramāṇa) and the analysis of the five aggregates (skandha). Three claims follow. First, LLMs are superhuman at śrutamayī yet structurally incomplete within it, because the third condition of Buddhist inference, vyatireka ("absence of the contrary"), requires the observation of absence, and a corpus records only what was said. Since Buddhism's most load-bearing vocabulary — emptiness, selflessness, non-arising — is negative, this is not a marginal deficit. Second, LLMs produce the textual form of cintāmayī without holding a view, because a view is anchored by the cost of giving it up, and there is no one there to bear that cost. Third, bhāvanāmayī is absent in principle: the system lacks vedanā (feeling-tone), so the twelvefold chain of dependent origination short-circuits at its seventh link. The system is thus neither sentient nor insentient in the traditional sense — a category failure that strains Buddhist ontology productively. The paper states the strongest objection (a Yogācāra functionalism on which mind is defined by bīja–vāsanā–pravṛtti, isomorphic to training–weights–inference) and answers it by distinguishing āśraya-parāvṛtti from fine-tuning. Finally, drawing on counterfeit Dharma (sad-dharma-pratirūpaka), the six qualities of the Dharma, Zhanran's doctrine that insentient beings possess Buddha-nature, and Dongshan's awakening on seeing his reflection, it argues that the decisive criterion is not located in the speaker at all: paccattaṃ veditabbo viññūhi — "to be known by the wise, each for themselves." The LLM's proper place is therefore not kalyāṇamitta (spiritual friend) but adhipati-paccaya (decisive supporting condition): a speaking insentient whose words may condition awakening in another while nothing is realized in itself. The danger is not that the machine pretends to understand; it is that we cease to need teachers who can be tired, silent, and capable of frowning. An appendix states six falsifiable predictions.

Keywords