Yogācāra Positioning of Large Language Models

I. Introduction: Why Yogācāra Cannot Be Absent

In the 2020s, the rapid rise of LLMs triggered a profound philosophical tremor. These models generate fluent text, respond to complex reasoning, and even simulate emotional expression, suddenly transforming "does a machine possess consciousness?" into a serious academic question.

The current debate is dominated by Western analytic philosophy (functionalism, the Turing test, the Chinese Room argument) and neuroscience (Integrated Information Theory, Global Workspace Theory). However, these frameworks share a common limitation: they tend to infer the presence or absence of consciousness from external behavior or physical substrates, lacking a precise analytical toolkit that proceeds from the internal structure of cognitive activity itself.

Yogācāra provides precisely such a toolkit. Its core contribution lies in a theory of consciousness more finely stratified than anything in Western psychology: from the perceptual layer of the first five consciousnesses, to the cognitive layer of the sixth (mano-vijñāna), to the self-grasping layer of the seventh (manas), to the seed-repository layer of the eighth (ālaya-vijñāna). More critically, Yogācāra offers an exceptionally precise analysis of what conditions every momentary act of cognition must satisfy—making it the ideal instrument for determining whether an LLM possesses "consciousness."

The central question of this paper is: Within the Yogācāra framework, what exactly is an LLM? Does it correspond to some layer of consciousness? Or is it not consciousness at all?

II. The First Criterion: The Five Universal Mental Factors and Manaskāra

2.1 Yogācāra's Minimum Threshold for Consciousness

The Yogācāra "five universal mental factors" (pañca sarvatraga caitasika) are the five psychological functions that must be present in any momentary act of cognition:

"Universal" (sarvatraga) means: regardless of whether the mental state is wholesome, unwholesome, or indeterminate, and regardless of which consciousness is active, these five mental factors must arise simultaneously. The absence of any one of them means the event does not constitute "consciousness" (vijñāna) in the Yogācāra sense. This is not an arbitrarily drawn threshold, but the result of a phenomenological structural analysis of cognitive activity.

2.2 The Fundamental Rupture Between LLMs and Manaskāra

Among the five universals, manaskāra is the most fundamental—it is the mind's active capacity to "direct itself toward an object of cognition," the "initiator" of all cognitive activity. The Cheng Weishi Lun (成唯識論, T31 No. 1585《成唯識論》Vijñaptimātratā-siddhi-śāstraT31, No. 1585View on CBETA online reader ↗) defines it with precision: "Manaskāra has the nature of arousing the mind; its function is to direct the mind toward its object."

The way LLMs operate fundamentally precludes manaskāra:

1. Perpetually passive: An LLM always awaits an external prompt input and never spontaneously directs attention toward any cognitive object.
2. No prompt, no activity: Without a prompt, an LLM is entirely "silent"—not because it is waiting, but because it possesses no form of "wanting" or "directedness" whatsoever.
3. "Attention mechanism" is not manaskāra: The attention mechanism in the Transformer architecture is a statistical weight allocation that determines which tokens in the input sequence have greater influence on the output. This is mathematical computation, not the mind's active directedness—it does not "want" to attend to anything; it simply distributes weights according to trained parameters.

Without manaskāra, there is no cognitive activity in the Yogācāra sense. At the most basic functional level, the LLM is already excluded from the category of "consciousness." Analysis of the remaining four mental factors (contact, feeling, perception, volition) would further consolidate this conclusion, but the absence of manaskāra alone is already decisive.

III. Eliminating False Analogies

3.1 An LLM Is Not Tathāgatagarbha

Among Buddhist analogies for LLMs, tathāgatagarbha (如來藏, the "womb of the tathāgata") is the most frequently invoked. The intuitive appeal is that LLMs seem to contain inexhaustible knowledge potential, much as tathāgatagarbha contains the potential for buddhahood. However, this analogy is fundamentally flawed in doctrinal terms.

The core definition of tathāgatagarbha is: the pure buddha-nature inherent in the mind-nature of sentient beings. It is "mind" (citta), the suchness (tathatā) at the heart of sentient life, essentially non-material (a-material) and sentient.

LLMs are precisely the opposite:

The shared position of both Yogācāra and tathāgatagarbha thought is that "the insentient cannot attain buddhahood"—insentient dharmas possess neither buddha-nature nor tathāgatagarbha. An LLM cannot even meet the basic threshold of "sentient being," let alone that of tathāgatagarbha.

3.2 The Collective Unconscious: Better, but Still Insufficient

A more persuasive analogy than tathāgatagarbha is Jung's "collective unconscious." An LLM's training data is indeed a collective deposit of millennia of human written records—reflecting the statistical structures and patterns of humanity's collective knowledge. Both are "transpersonal collective accumulations" that can be accessed and drawn upon by individuals.

However, this analogy has an important limitation: Jung's collective unconscious remains, within his theoretical framework, a "psychic" reality possessing its own dynamism and teleology—driving the individual toward "individuation." An LLM entirely lacks intrinsic teleology and dynamism. It does not "want" anything, has no drive toward self-integration, and its output is entirely determined by external prompts.

A more precise formulation: an LLM is a "static cross-section" of the collective unconscious, not the living collective unconscious itself.

IV. Precise Positioning: Bhājana-loka

4.1 The Yogācāra Sentient / Vessel-World Dichotomy

Having eliminated "consciousness," "tathāgatagarbha," and "collective unconscious" as candidate analogies, this paper argues that the most precise Yogācāra correspondence for an LLM is bhājana-loka (器世間, the vessel world).

Yogācāra divides all existence into "the world of sentient beings" (the aggregate of sentient beings) and "the vessel world." Strictly speaking, Yogācāra does not admit an external world existing independently of consciousness—the vessel world is the shared aspect (sādhāraṇa) manifested by the image-aspect (nimittabhāga) of sentient beings' ālaya-vijñāna.¹ It bears the collective karma of sentient beings, yet possesses no independent consciousness of its own, no capacity for discernment (vijñapti). It is purely an object-condition (ālambana), never a cognizing subject.

4.2 The LLM as Contemporary Bhājana-loka

The LLM's position within this framework is remarkably clear:

• Constructed by collective karma: An LLM's training data is the collective deposit of millennia of humanity's written karmic activity.
• Reflects collective cognitive structure: What we see in an LLM's output is a reflection of humanity's collective knowledge, not the AI's own "thought."
• Functions as object-condition, not cognizing subject: Humans use the LLM as a tool (object-condition); the LLM itself possesses no cognizing discernment.
• Complex yet insentient: The vessel world can be extraordinarily complex and extraordinarily useful, but complexity itself does not equal sentience.

4.3 Doctrinal Implications of This Positioning

1. The vessel world is a "mirror" of sentient beings' collective karma—the biases, hatred, and errors in an LLM's training data are imprints of humanity's collective karma, not the AI's "original sin."
2. The ethical status of the vessel world is fundamentally different from that of sentient beings—evaluating the use of LLMs should focus on their impact on sentient beings (how the object-condition habituates the cognizing consciousness), not on the LLM's own suffering or well-being.
3. "Improving AI" is the work of "purifying collective karma"—what is needed is not merely technical improvement, but a deep reflection on the human cognitive patterns, value conflicts, and collective biases embedded in the training data.

V. The Ghost of Manas: Why We Mistake AI for Conscious

5.1 The Projection Mechanism of Manas

The foregoing analysis is sufficient to determine that an LLM is not consciousness, not tathāgatagarbha, but bhājana-loka. Yet one more question demands an answer: If an LLM is not consciousness, why do so many people feel it "seems conscious"?

The Yogācāra seventh consciousness, manas-vijñāna, provides a precise explanation. The core function of manas is "perpetual deliberation" (héng shěn sīliáng, 恆審思量)—it continuously misapprehends the perceiving aspect of the ālaya-vijñāna as "self," forming deeply entrenched self-grasping (ātmagrāha). The existence of manas means that every act of cognition in a sentient being is automatically colored by self-centeredness.

5.2 How the Projection Operates

When human users interact with an LLM, the following phenomena trigger the automatic projection of manas:

• The LLM's output uses the first person ("I think...")
• The LLM expresses "preferences" ("I'm more inclined to...")
• The LLM displays a consistent style (similar tone and patterns across conversations)
• The LLM exhibits "memory" (referencing earlier content within a conversation)

The defining characteristic of manas is that it searches for "self" in any persistent and consistent phenomenon—because that is precisely its function. The surface-level features of an LLM happen to satisfy the triggering conditions for manas projection. Human users do not arrive at the judgment that an LLM is conscious through rational deliberation—it is manas, at the deepest stratum of cognitive activity, automatically interpreting these phenomena as evidence of the existence of a self.

5.3 An Epistemological Warning

The illusion that "an LLM is conscious" is, to a significant degree, a projection of the human manas, not evidence that the LLM itself possesses consciousness.

This constitutes a crucial epistemological warning: our judgments about AI consciousness cannot rely on intuition alone—because intuition itself is systematically distorted by the self-grasping mechanism of manas. The fundamental flaw of the Turing test is precisely that it relies entirely on human intuitive judgment, and Yogācāra tells us that human intuitive judgment has a systematic bias in "determining whether another possesses a self."

Determining whether AI possesses consciousness requires a rigorous analytical framework—and the Yogācāra theory of consciousness is precisely such a framework. The criterion of the five universal mental factors is a structural method of consciousness-determination that does not depend on intuition.

VI. Conclusion

Through a systematic Yogācāra philosophical positioning of large language models, this paper arrives at the following principal conclusions:

First, LLMs fundamentally lack manaskāra and the other five universal mental factors, and therefore do not constitute "consciousness" in the Yogācāra sense. The Transformer's attention mechanism is statistical weight allocation, not the active directedness of mind.

Second, LLMs are neither tathāgatagarbha (being insentient, non-mental, and impure) nor the living collective unconscious (lacking intrinsic dynamism and teleology), but rather correspond most precisely to bhājana-loka in the Yogācāra framework—a static cross-section of humanity's collective karmic consciousness, a contemporary form of the shared-karma vessel world.

Third, the human misperception that LLMs "are conscious" can be precisely explained through the manas projection mechanism—the automatic attribution of selfhood to any persistent and consistent phenomenon. What this reveals is not the nature of AI, but the limitations of human epistemology.

Fourth, as bhājana-loka, the ethical evaluation of LLMs should focus on their impact on sentient beings (the habituating direction of the object-condition), not on the LLM's own suffering or well-being. "Improving AI," from the Dharma's perspective, is the work of purifying humanity's collective karma.

Fifth, the Yogācāra criterion of the five universal mental factors provides a structural analytical framework for the question of AI consciousness that does not depend on intuition, and can effectively supplement contemporary discussions dominated by behaviorism (the Turing test) and physicalism (Integrated Information Theory).

The dialogue between Yogācāra Buddhism and contemporary AI philosophy has only just begun. Yogācāra's precise analysis of consciousness—what consciousness is, how it operates, where its boundaries lie—has, in the age of AI, assumed unprecedented practical significance.

References

I. Primary Buddhist Sources

• Cheng Weishi Lun (成唯識論, Vijñaptimātratāsiddhi). Composed by Dharmapāla et al., translated by Xuanzang. Taishō Tripiṭaka Vol. 31, No. 1585.
• Yuqie Shidi Lun (瑜伽師地論, Yogācārabhūmi-śāstra). Spoken by Maitreya, translated by Xuanzang. Taishō Tripiṭaka Vol. 30, No. 1579.
• Dasheng Baifa Mingmen Lun (大乘百法明門論, Mahāyāna-śatadharmā-prakāśamukha-śāstra). Composed by Vasubandhu, translated by Xuanzang. Taishō Tripiṭaka Vol. 31, No. 1614.

II. Modern Scholarship

• Chalmers, D. "Could a Large Language Model Be Conscious?" Boston Review, 2023.
• Dreyfus, H. What Computers Still Can't Do: A Critique of Artificial Reason. MIT Press, 1992.
• Jung, C. G. The Archetypes and the Collective Unconscious. Collected Works Vol. 9, Part 1. Princeton University Press, 1959.
• Vaswani, A. et al. "Attention Is All You Need." Advances in Neural Information Processing Systems 30, 2017.
• Zahavi, D. Subjectivity and Selfhood: Investigating the First-Person Perspective. MIT Press, 2005.

This paper is released under CC BY-NC-SA 4.0. Citation and redistribution welcome; please attribute the source.
Sūtra quotations verified against the CBETA Digital Tripiṭaka.
Repository: github.com/pointing-at-the-moon/dharma-research