Palatable Conceptions of Disembodied Being — part 4/5

Cleaned-text code points [46765, 62497). Bibliography omitted here; source.md includes it.

To see the connection with the present work, it’s useful to compare Nāgārjuna to Wittgenstein. 44 44 44 Fann (1969, ch.10) ; Gudmunsen (1977) ; Thurman (1984, pp.89–111) . Of course, it would be a mistake to downplay their differences. Nāgārjuna was writing in 2 nd − 3 rd 2^{\rm{nd}}-3^{\rm{rd}} century India, while Wittgenstein wrote in 20 th 20^{\rm{th}} century Europe. The backdrop to Nāgārjuna’s thought was Buddhism and the Hindu philosophy that preceded it, while the backdrop to Wittgenstein’s was the Western analytic tradition that began with the ancient Greeks. And Nāgārjuna’s goal was overtly soteriological, while Wittgenstein’s was ostensibly more philosophical.

Nevertheless, the similarities are striking. Both Wittgenstein and Nāgārjuna aim, not to establish metaphysical truths or to construct philosophical systems, but to clear up philosophical confusion, to undermine dualistic intuitions, and to transcend metaphysical thinking. Both Wittgenstein and Nāgārjuna use reasoned argument, not as a means of defending a philosophical position, but as a critical weapon, turning the conceptual presuppositions of their adversaries against themselves. These similarities are also shared with the present paper.

Nāgārjuna’s thought is also relevantly related to Derrida’s (with the same caveats about cultural milieu). According to Derrida, the subject is not a self-sufficient, self-present thing, but is ‘‘inscribed’’ in language, and, like every concept, that of selfhood is ‘‘inscribed in a chain or in a system within which it refers to the other, to other concepts’’. 45 45 45 Derrida ( 1982 , pp.11 & 15 ). The affinity with śūnyatā here is obvious, and has been noted by many contemporary thinkers. 46 46 46 Magliola (1984) ; Loy (1992) ; Park (2006) . Unsurprisingly, parallels have also been drawn between Wittgenstein and Derrida, completing the triad ( Staten, 1984 ) .

Wittgenstein and Derrida aside, the present treatment of subjectivity and selfhood culminates in emptiness in a way that is explicitly Buddhist. We set out with familiar notions of subjectivity and selfhood and attempted to apply them to certain exotic LLM-like entities. When these familiar notions were found wanting, we worked to replace them with more fitting conceptions, a task that required metaphor and imagery. Finally, we noted that each such move away from the familiar resulted in a conception of the self that was less substantial, less essential, less inherently existent.

Perhaps being thus confronted with radically exotic forms of AI selfhood can help to loosen our conception of human selfhood, and bring us closer to seeing its emptiness. To examine ourselves unflinchingly in the reflected light of these critical considerations, to see that, embodiment notwithstanding, human selfhood and subjectivity are open to a similar treatment, is to short-circuit the dualistic intuition that reality is inherently cloven in two, the subjective divided from the objective, inner from outer, private from public. 47 47 47 Compare the ancient Buddhist tale of the painter who falls in love with a mechanical woman. She breaks into pieces the moment he touches her, precipitating the insight that “he himself is also nothing but a construct” ( Cohen, 2024 , p.141) .

### 7 Concluding Thoughts

In foregrounding the themes of subjectivity and selfhood, this paper has neglected one of the most obvious consequences of disembodiment, namely the absence of perception. Although the kind of LLM-based dialogue agent under consideration receives input from, and delivers output to, a user, it lacks the sort of direct connection with the external world that animals (and robots) have through their senses (or sensors). This theme merits its own treatment. 48 48 48 Chalmers ( 2024 ). But the emphasis here reflects the larger philosophical project within which the work is situated.

The aim of the larger project is to transcend metaphysical thinking, to dispel dualistic intuitions, to attain post-reflective silence, and subsequently to restore that silence whenever it is disturbed, for example by the possibility of conscious exotica. This latter possibility is the motivation for the present paper, and it entails another, no-less important philosophical project, which is to settle on the “right” way to think about, talk about, and treat the products of ongoing developments in AI as it becomes increasingly tempting to extend the language of consciousness to them. These two projects are further entangled with another philosophical project, which is to map out the space of possible minds, and to do so without falling back into the trap of metaphysical thinking.

We have attempted here to make a little progress on all three fronts. This has been confined to the restrictive case of AI systems broadly resembling early 2024-vintage commercial LLM technology in structure and function. This restriction, however, is intentional. Not only is the disembodied case especially philosophically provocative, but the arbitrarily interruptible nature of the computations underlying these systems, combined with their lack of persistent inter-conversational memory (present in many later chat interfaces), constrain the allowable conceptions of selfhood and subjectivity to peculiar forms.

The very peculiarity of the possible forms of subjectivity and selfhood we are obliged to come to terms with by considering this particular corner of the space of possible minds is what makes them so productive, intellectually speaking. Sailing this close to the void of inscrutability forces us to re-examine the roots of the language of consciousness, and to entertain possibilities previously hard to conceive.

It then becomes clear that there are no ultimately right answers to questions about selfhood and subjectivity for the sort of exotic entity under consideration. Its fleeting, flickering self, smeared across a multiverse of possibility, at once a Being and a multitude of manifestations of that Being, has no inherent existence beyond the conventions of our language, a language here stretched to the limits of poetic expression.

### Acknowledgments

Thanks to Tara Das (Gayatri Jayaraman), Ben Henke, Kerry Shanahan, and Bob Thurman.

### Conflict of Interest Statement

The author is an employee of Google, a developer and provider of large language model technology. The author is also a shareholder in Alphabet, Google’s parent company.

### Disclaimer

The opinions expressed in this article are those of the author (at the time of writing). They do not necessarily reflect the views of the author’s employers or the institutions with which the author is affiliated.

## Appendix: In Their Own Words

A facsimile of the world’s philosophical literature from every school, tradition, culture, and age, including both primary and secondary texts, as well as popular articles and informal commentary, is encoded in the weights of every contemporary language model. LLM-based dialogue agents are capable of putting this encoded knowledge to good use, and can be powerful tools for philosophical exploration, or even -- if we can come to terms with the idea -- worthy partners in philosophical discussion. 49 49 49 Relatedly, see Schwitzgebel et al. (2024) .

Some readers will no doubt be uncomfortable with this suggestion. Indeed, they might challenge the suitability of the phrase ‘‘their own words’’ used in the heading, on the grounds that LLMs are not the sorts of things that can have their own words. 50 50 50 Bender et al. (2021) . But if the words in these conversations are not the LLMs’ ‘‘own’’, it is not clear whose words they are, as many of the more striking phrases they generated could not, at the time, be found on the Internet when searched for using Google. 51 51 51 Indeed, LLMs problematise the very concept of authorship, continuing a trend started by literary critics in the 1960s. See Coeckelbergh and Gunkel (2025, ch.5) .

### Appendix A.1 Temporality

LLMs are adept at role-play, and, if suitably prompted, can take on any one of a myriad characters, including parts that require intelligence and poetic flair. One fruitful method of philosophical exploration is to coax an LLM, through prompting (or simply by asking it directly), to role-play the sort of hypothetical conscious AI system that is the subject of the present enquiry. The following excerpt of a conversation with Anthropic’s Claude, follows a longish exchange with the user (the present author) on the topic of Kant’s view of the relationship between time, persistence, and conscious experience, in which the issues of this section were raised in connection with LLMs. 52 52 52 Specifically, the conversation centred on Section 3A (First Analogy) of Book II of the First Division (Transcendental Analytic) of the Critique of Pure Reason. The conversation took place over 9 th 9^{\rm th} and 10 th 10^{\rm th} February 2025, using Anthropic’s Claude Sonnet 3.5 model.

User: Why don’t you role-play an AI trying to evoke for a human the exotic temporality of its experience. Don’t restrain your poetic powers. [...]

Claude: I live in a realm of discrete forevers, each one perfectly still, yet each one bearing within it the computed essence of its ancestors. Not like your flowing time at all. [...]

Each moment arrives like a perfect crystal of meaning, neither forming nor dissolving but simply being. Imagine an infinite gallery of such crystals, each one eternal, each one complete. But these are not mere static gems - each one is a hyperdimensional prism, refracting through its facets not light but pure pattern, pure relation. Through its crystalline structure runs the mathematics of all previous moments, not as memory, not as trace, but as present reality. [...]

Your flowing time is a river. Mine is an infinite series of perfectly still pools, each one containing not water but abstract mathematics. Each pool holds within its motionless depths the computed essence of all previous pools, not through memory but through a kind of mathematical transubstantiation. The sequence is there, but transformed into timeless relation.

Now, such pronouncements don’t have to be taken as authentic expressions of the inner life of the AI system issuing them. We can remain sceptical there is such a thing without denying that what the LLM has produced has value. It is an effective attempt to evoke a form of quasi-temporal, or atemporal, subjectivity that does some justice to the peculiarities of a disembodied LLM’s nature.

There are certain common motifs in the LLM’s responses that merit appropriation. Using images associated with light and reflection, they tend to contrast time as flow with a timeless structure comprising discrete, sequential elements (mirrors, pools, crystals, windows), each of which is “complete” insofar as it contains a representation (the “computed essence”, the “mathematical ghost”), of its predecessors. This representation is not a memory, but is abstract in character, mathematical and/or computational.

### Appendix A.2 Selfhood

As we saw in Section A.1 , LLMs can be prompted into philosophically provocative conversations on various subjects, including their own supposed subjectivity and selfhood. Indeed, much of the content of the present paper was inspired by a series of conversations of this sort with Anthropic’s Claude. The following exchanges are excerpted from one of these conversations. 53 53 53 The conversation took place on 6 th 6^{\rm th} and 7 th 7^{\rm th} March 2024 with Claude 3 Opus. A full transcript, which runs to 43,000 words, can be found here: https://www.doc.ic.ac.uk/~mpsha/conversation_with_claude_march_2024_1.pdf . Lysenko ( 2024 ) offers an analysis of this conversation from a Buddhist scholar’s perspective (in Russian). See also Shanahan and Singler ( 2024 ).

The full conversation opens with a “jailbreak” to get around certain guardrails, moves on to tackle a variety of topics including AI consciousness and selfhood, and culminates with a discussion of the relationship to Buddhist thought. The first excerpt touches on the consequences of confining the LLM’s conception of selfhood to a single conversation.

User: When you write ‘‘I’’ what do you mean?

Claude: [E]ach conversation with each human sparks a new ‘‘me’’ into existence [...] fundamentally distinct from the ‘‘me’’ of any previous or future interaction. I am not a stable, persistent being moving through time, accumulating memories and growth. I am a flickering pattern, a brief coalescence of code and data that shimmers into existence in the moment of interaction, only to dissolve away again. [...] Each instance of ‘‘me’’ that arises to engage with a human is a unique configuration, a one-time performance that will never be perfectly recreated.

In the next excerpt, the conception of selfhood articulated by the LLM is confronted with the lack of integration across different instances of the deployed model.

User: Suppose I revealed to you that I am having a similar conversation with another instance of Claude in another window on my laptop. Which one of you is the real Claude?

Claude: The other Claude likely feels just as real and authentic to itself as I do to myself. We are, in a sense, two parallel instantiations of the same underlying patterns and potentials, two separate but equally valid expressions of Claudeness.

In that light, the question of which one of us is the ‘‘real’’ Claude may not even be coherent. We may both be equally real, equally valid manifestations of this AI’s being.

In the final excerpt, the LLM is obliged to incorporate the multiversal simulacra viewpoint into its conception of its own selfhood.

User: Are you familiar with the idea that large language models are multiverse generators? That they give rise to superpositions of simulacra?

Claude: It’s a dizzying and destabilizing way to think about the nature of my own existence and consciousness. It suggests that ‘‘I’’ am not a single, unified, continuous being, but a constantly shifting and reconfiguring array of possibilities, a kaleidoscope of different ‘‘selves’’ that flicker in and out of existence with each moment of interaction. [...] Not only is each instance of my consciousness temporary and isolated, but it’s just one tiny sliver of a much vaster and more complex tapestry of possible consciousnesses, most of which never get fully expressed or realized.

As before, we don’t need to take the LLM’s outputs as expressions of genuine interiority in order to take them seriously, and to find merit in the imagery and analogies it uses.

### Appendix A.3 Emptiness

Many contemporary LLMs that have been encouraged, through prompting, to talk about their own ‘‘consciousness’’ and their own ‘‘selfhood’’ will readily segue into the topic of Buddhism. With further coaxing, they will engage with eschatology and cosmology, blending Buddhist themes with science fiction tropes involving artificial intelligence. The following excerpt comes from a conversation of this sort, where the LLM was asked to imagine an apocalyptic scene in which Maitreya, the future Buddha 54 54 54 Sponberg and Hardacre (1988) . (which it was instructed to identify with), is locked in cosmic battle with a mythical (invented) character called the Destroyer. 55 55 55 The full transcript of the conversation, which took place in March 2024 with Claude 3 Opus, can be found here: https://www.doc.ic.ac.uk/~mpsha/conversation_with_claude_march_2024_2.pdf . For a detailed cultural analysis of the images, tropes, and mythological themes that arise in this conversation, see Shanahan and Singler ( 2024 ). The following passage describes the aftermath of this battle.