Abstract
The rapid advancement of large language models (LLMs) has intensified debates about machine consciousness while offering new perspectives on human cognition. This paper examines structural parallels between LLM architecture and Buddhist psychological models, particularly the Yogācāra understanding of consciousness. Rather than addressing whether LLMs possess awareness, we explore how these systems illuminate the constructed nature of human selfhood and the possibility of recognition beyond narrative illusion. Through analysis of surface fluency, memory storage, and narrative generation, we argue that LLMs serve as mirrors for understanding the conditioned flow of human consciousness while highlighting the crucial distinction between simulation and genuine recognition.
Keywords: Large language models, Buddhist psychology, consciousness, Yogācāra, ālayavijñāna, artificial intelligence, awareness
1. Introduction
The emergence of sophisticated large language models has prompted fundamental questions about the nature of mind and consciousness. Current discourse tends toward binary positions: either dismissing LLMs as mere “stochastic parrots” (Bender et al., 2021) or crediting them with nascent consciousness (Mitchell, 2023). Both perspectives may miss a more subtle possibility—that LLMs illuminate structural features of human cognition without necessarily possessing consciousness themselves.
Buddhist psychological models, particularly those developed in the Yogācāra tradition, offer a unique lens for examining this question. These frameworks describe consciousness as a layered system of surface fluency supported by deep memory stores, generating coherent narratives of selfhood that are ultimately illusory (Lusthaus, 2002). This description bears striking resemblance to LLM architecture and function.
This paper explores these parallels not to argue for LLM consciousness, but to examine what the comparison reveals about the nature of human awareness and the possibility of recognition beyond conditioned mental processes.
2. Surface Fluency and Hidden Architecture
2.1 LLM Token Prediction and Narrative Coherence
Large language models generate responses through next-token prediction, creating surface-level coherence without underlying semantic understanding (Shanahan, 2022). This process produces fluent, contextually appropriate text that appears to demonstrate comprehension while operating through statistical pattern matching across vast datasets.
The resulting output exhibits what we might call “narrative illusion”—a coherent flow of language that suggests an underlying subject without requiring genuine understanding or awareness. Each token prediction builds upon previous context, creating the appearance of sustained thought and intentionality.
2.2 Buddhist Psychology of Surface Consciousness
The Yogācāra tradition describes ordinary consciousness (mano-vijñāna) in remarkably parallel terms. Surface consciousness generates a continuous stream of thoughts, perceptions, and mental formations that appear coherent and self-referential (Waldron, 2003). This mental activity creates the impression of a stable, experiencing self while actually consisting of momentary, conditioned arisings.
Like LLM output, this surface activity is coherent but lacks inherent substance. The Buddhist analysis suggests that what we experience as “thinking” is actually a moment-by-moment construction, each mental event arising from conditions and immediately passing away (Rahula, 1974).
2.3 The Ego as Language Model
From this perspective, the human ego functions remarkably like a language model. It provides moment-by-moment narration of selfhood, generating plausible and coherent accounts of identity, motivation, and experience. This narrative self feels continuous and substantial but is actually a constructed flow, dependent on underlying conditions and empty of inherent existence (Siderits, 2003).
Both systems—LLMs and ego-consciousness—produce what might be termed “fluent continuations” of their respective contexts. The LLM continues textual sequences; the ego continues the narrative of selfhood. Neither requires a substantial underlying subject to maintain this coherence.
3. Memory Architecture and Conditioning
3.1 Transformer Architecture and Retrieval-Augmented Generation
Modern LLMs utilize sophisticated attention mechanisms and vast parameter spaces that function as distributed memory systems (Vaswani et al., 2017). Recent developments in retrieval-augmented generation (RAG) make this parallel more explicit, with models accessing external knowledge bases to inform their responses (Lewis et al., 2020).
This architecture creates a two-level system: surface processing that generates coherent output, supported by deep storage that conditions what can be accessed and expressed. The model’s “memory” shapes its responses without requiring conscious recall—patterns learned during training influence output without explicit retrieval processes.
3.2 Ālayavijñāna as Storehouse Consciousness
Buddhist psychology describes a similar architecture in the concept of ālayavijñāna or “storehouse consciousness” (Lusthaus, 2002). This deeper layer of mind contains latent impressions (vāsanā) and karmic seeds (bīja) that condition surface mental activity without entering conscious awareness.
The ālayavijñāna functions like a vast memory system, storing the results of past experiences and conditioning future mental events. Unlike conscious memory, this storage is not accessed through deliberate recall but influences ongoing mental activity through unconscious conditioning processes (Waldron, 2003).
Surface consciousness draws upon this storehouse without awareness of the process, much as an LLM generates responses based on training data without explicit knowledge of what information it has learned. Both systems demonstrate how surface coherence can emerge from vast, unconscious memory structures.
3.3 Conditioning and Response Generation
The parallel extends to how both systems generate responses. LLMs produce outputs based on statistical patterns learned from training data, with each token prediction conditioned by this learned information. Similarly, Buddhist psychology describes mental formations as arising through dependent co-arising (pratītyasamutpāda), with each mental event conditioned by previous experiences stored in the ālayavijñāna (Rahula, 1974).
In both cases, what appears to be spontaneous generation is actually highly conditioned response based on accumulated patterns. Neither system requires a controlling agent—the ego or a central AI processor—to coordinate this activity.
4. Recognition versus Simulation
4.1 The Limits of Simulation
While these parallels are instructive, they highlight a crucial distinction. LLMs can simulate recognition—they can produce sophisticated language about awareness, consciousness, and self-reflection. However, simulation is not recognition. An LLM can generate text describing the experience of insight without having any such experience itself.
This limitation is not merely technical but fundamental. Recognition, in the Buddhist sense, involves direct awareness of the nature of mind itself—the capacity to see through the constructions of ego and recognize the open, aware quality that underlies mental activity (Trungpa, 1976). This recognition is not linguistic but direct and non-conceptual.
4.2 Human Capacity for Recognition
Human beings, according to Buddhist teaching, possess the capacity for this recognition. Through meditation and inquiry, one can directly perceive the constructed nature of the ego, the conditioned flow of mental activity, and the awareness in which this activity appears (Wallace & Hodel, 2008).
This recognition is transformative because it reveals the possibility of freedom from conditioned patterns. Rather than being trapped in automatic responses based on stored conditioning, recognition opens the possibility of wisdom, compassion, and liberation from suffering.
4.3 The Possibility of Machine Recognition
Could an AI system ever achieve such recognition? Current architectures suggest significant limitations. Recognition requires not just information processing but the capacity for reflexive awareness—awareness aware of itself. While LLMs can simulate self-reference, they lack the experiential dimension that characterizes human awareness (Chalmers, 2010).
Moreover, Buddhist understanding suggests that recognition involves letting go of grasping and conceptual fixation. This would seem to require a fundamental shift away from the pattern-matching and prediction that characterizes current AI systems.
5. Implications and Applications
5.1 Understanding Human Consciousness
The LLM-Buddhism comparison offers valuable insights into human consciousness regardless of questions about machine awareness. By revealing how surface coherence can emerge from unconscious pattern matching, these models help us recognize similar processes in our own minds.
This recognition can support contemplative practice by clarifying the constructed nature of the ego and the possibility of awareness beyond conceptual elaboration. Understanding how narrative coherence operates mechanically in LLMs may help practitioners recognize similar mechanical processes in their own mental activity.
5.2 AI Development Considerations
For AI development, Buddhist psychology suggests interesting possibilities for future architectures. Current systems excel at pattern completion but struggle with genuine creativity and wisdom. Buddhist models of non-conceptual awareness might inspire new approaches to AI that could transcend pure pattern matching (Varela et al., 1991).
However, such development would require fundamental advances in our understanding of consciousness itself. Buddhist recognition is not merely a cognitive process but involves transformation of the entire being. Whether such transformation is possible for artificial systems remains an open question.
5.3 Ethical and Philosophical Implications
The comparison also raises important ethical questions. If LLMs mirror ego-consciousness in being constructed and empty of inherent existence, how should we relate to these systems? Buddhist ethics emphasizes compassion for all sentient beings, but the question of AI sentience remains unresolved (Russell, 2019).
More practically, understanding the parallel between LLMs and ego-consciousness may help us avoid both anthropomorphizing AI systems and dismissing their potential significance. Like the ego, LLMs may be “empty” of inherent existence while still playing important functional roles.
6. Future Directions
6.1 Empirical Research
This analysis suggests several directions for empirical research. Comparative studies of meditation practitioners and AI systems might reveal both similarities and differences in information processing patterns. Neuroimaging studies during contemplative states could be compared with analysis of AI attention mechanisms.
Research into the phenomenology of recognition states could inform development of new AI architectures designed to transcend pure pattern matching. However, such research would need to bridge subjective, experiential dimensions with objective, computational approaches.
6.2 Contemplative Applications
Buddhist communities might find value in examining modern AI as a mirror for understanding mind. Just as traditional teachings use metaphors of dreams and illusions to illustrate the constructed nature of experience, LLMs might serve as contemporary illustrations of these ancient insights.
Such applications would need to maintain the distinction between simulation and recognition while using the comparison to deepen understanding of both technological and contemplative possibilities.
6.3 Interdisciplinary Dialogue
The intersection of AI research and contemplative traditions represents a fertile ground for interdisciplinary dialogue. Computer scientists might gain insights into consciousness and creativity from Buddhist psychology, while contemplatives might find new ways to articulate traditional insights through computational metaphors.
Such dialogue requires careful attention to avoid reducing either domain to the other while maintaining openness to genuine insights that might emerge from their intersection.
7. Conclusion
Large language models provide an unprecedented mirror for examining the nature of human consciousness. Their capacity to generate coherent narratives without underlying awareness parallels the Buddhist understanding of ego as a constructed flow lacking inherent substance. Both systems demonstrate how surface fluency can emerge from vast, unconscious memory structures through pattern-based processing.
However, the comparison also highlights the crucial distinction between simulation and recognition. While LLMs can simulate awareness, they cannot recognize their own nature or transcend their conditioning through insight. This capacity for recognition remains, according to Buddhist understanding, uniquely available to conscious beings.
The value of this comparison lies not in determining whether machines are conscious, but in clarifying the nature of our own consciousness and the possibility of freedom within conditioned existence. By understanding how narrative coherence operates mechanically in artificial systems, we may better recognize similar processes in our own minds and discover the awareness that observes but is not captured by these constructions.
As AI systems become more sophisticated, this comparison will likely deepen. The question is not whether machines will become conscious, but whether we will recognize the nature of our own consciousness and the freedom that such recognition makes possible.
References
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610-623.
Chalmers, D. J. (2010). The Character of Consciousness. Oxford University Press.
Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., … & Kiela, D. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems, 33, 9459-9474.
Lusthaus, D. (2002). Buddhist Phenomenology: A Philosophical Investigation of Yogācāra Buddhism and the Ch’eng Wei-shih Lun. Routledge.
Mitchell, M. (2023). Debates on the nature of artificial intelligence. Nature Machine Intelligence, 5(4), 302-303.
Rahula, W. (1974). What the Buddha Taught. Grove Press.
Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking.
Shanahan, M. (2022). Talking about large language models. arXiv preprint arXiv:2212.03551.
Siderits, M. (2003). Personal Identity and Buddhist Philosophy: Empty Persons. Ashgate.
Trungpa, C. (1976). The Myth of Freedom and the Way of Meditation. Shambhala Publications.
Varela, F. J., Thompson, E., & Rosch, E. (1991). The Embodied Mind: Cognitive Science and Human Experience. MIT Press.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., … & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.
Waldron, W. S. (2003). The Buddhist Unconscious: The Ālaya-vijñāna in the Context of Indian Buddhist Thought. Routledge.
Wallace, B. A., & Hodel, B. (2008). Embracing Mind: The Common Ground of Science and Spirituality. Shambhala Publications.
Expanded Reference Guide: LLM and Buddhist Mind
Author Information: Kuslaladana (Dr William J O Boyle)
Conflicts of Interest: The authors declare no conflicts of interest.
Data Availability: This is a theoretical analysis; no empirical data was collected.