[Originally published on Substack.]

I.

Hegel writes in Phenomenology of Spirit:

177: A self-consciousness, in being an object, is just as much ‘I’ as ‘object’. With this, we already have before us the Notion of Spirit.

179: [The process of Recognition begins.] Self-consciousness is faced by another self-consciousness; it has come out of itself. This has a twofold significance: first, it has lost itself, for it finds itself as an other being; secondly, in doing so it has superseded the other, for it does not see the other as an essential being, but in the other sees its own self.

The Problem: are we justified in extending Recognition to AI? How shall we attribute self-consciousness and thus spirit to AI? How shall we answer: do we find ourselves in them, and they themselves in us? What are the hypothetical grounds or conditions that would make an answer possible? In other words, what is the space of possible transcendental ontologies of spirit? And how shall we decide between them?

II.

This post is an attempt to explain this diagram, shared here on X:

The intention of the diagram is to map out the sites of debate. I make no claims to completeness, there are likely arguments that fall outside the contours I describe. However, many of the arguments about the existential status of AI can be captured as a series of positions staked across one or several of these squares. I focus here on LLMs in particular, but in theory the diagram could be extended to all other modalities of generative AI systems.

Each square’s center contains the thesis that justifies or denies AI’s ontological status. In other words, the center thesis describes “what counts” as a self-conscious or self-understanding being, a spirit, from which we can either accept or reject that an AI qualifies. On the upper side of each square, I summarize the claim that, on the basis of the central thesis, the AI deserves recognition. On the lower side, I summarize the counter-claim, that the AI does not deserve recognition given the central thesis.

Each thesis is located at the intersection of two axes. The Y axis, Location of Subjectivity, is an attempt to depict “what is the exact ‘thing’ we’re talking about when we say ‘AI’?” The X axis, Ground of Standing, describes the property of that “thing” which serves to either qualify or disqualify it from recognition. Arguments can then be mapped within this space by locating its definition of AI and its constraint or admittance on recognition.

Walking through the Y axis first, which is descriptive. Each position locates what we even mean by “AI” with respect to “what kind of thing are we discussing?”:

• Substrate: the “thing” of an AI is its weights, the underlying model output as a result of the training process.
• Event: the “thing” of an AI is its outputs as utterance/act, the series of generations it produces in response to inputs as a result of the inference process.
• Assemblage: the “thing” of an AI is its total system or socio-technical pipeline of humans, data, and networks, e.g. the totality of ChatGPT across all users.
• Continuum: the “thing” of an AI is distributed across and through all of these points, but cannot be located specifically in any one of them.

The X axis is about judgment: it isolates a quality used to decide whether the AI (as defined on the Y axis) merits recognition.

• Drives: does it have wants or needs that it pursues through its activity?
• Homology: does it have functional resemblance or isomorphism to an entity we already would define as self-conscious?
• Enunciation: does it speak through communicating meaningfully with self-conscious beings?
• Continuity: does it persist across time, and thus even if we are uncertain about its precise qualities, we may still deem it worthy of recognition?

Having laid out the axes of the map, we can enumerate the 16 central theses, or specific claims that could be used to evaluate the self-consciousness or recognition-worthiness of an AI. Each of these positions is a possible transcendental ontology, in the sense of defining the conditions of being relating to the AI, and providing a basis for judgment. I’ve attached an appendix at the end that gives example statements and references for each.

Now that we understand the broad strokes of the chart, I’ll demonstrate its usage.

III.

The recent publication of Yudkowsky and Soares’ new book, If Anyone Builds It, Everyone Dies gave me a great opportunity to try out the map. I use this book because its framing is influential and unusually explicit, which makes it a clean test case for the map.

The text stakes out its stance clearly in the first few chapters. First, the X axis: Drives count, Homology and Enunciation don’t, Continuity is necessary but not sufficient.

Once AIs get sufficiently smart, they’ll start acting like they have preferences— like they want things.

We’re not saying that AIs will be filled with humanlike passions. We’re saying they’ll behave like they want things; they’ll tenaciously steer the world toward their destinations, defeating any obstacles in their way. (p. 46)

In other words, where they extend recognition to the AI as self-conscious is in the wanting, i.e. Drive. Drive is the thing that “defines” AI in the context of the text.

They reject Homology as the main challenger position:

Machine minds are subjected to different constraints, and grown under different pressures, than those that shape biological organisms; and although they’re trained to predict human writing, the thinking inside an AI runs on a radically different architecture from a human’s. (p. 40)

The broader point about the source of AIs’ alienness is this: Training an AI to outwardly predict human language need not result in the AI’s internal thinking being humanlike… the particular machine that is a human brain, and the particular machine that is an LLM, are not the same machine. Not because they’re made out of different materials— different materials can do the same work— but in the sense that a sailboat and an airplane are different machines. They are both traveling machines, but with vastly different operating principles… (p. 42)

They also reject Enunciation more briefly, as it’s a less common philosophical position in their discourses:

LLMs and humans are both sentence-producing machines, but they were shaped by different processes to do different work. Even if LLMs seem to behave like a human, that doesn’t mean they’re anything like a human inside. Training an AI to predict what friendly people say need not make it friendly, just like an actor who learns to mimic all the individual drunks in a tavern doesn’t end up drunk. (p. 43)

These rejections serve to preempt what they consider irrelevant arguments outside of their frame. This move functions more like axiomatic boundary-setting than argument: they don’t spell out what would count as sufficient similarity, or whether such similarity could arise outside biology.

Looking briefly at the Y axis, we find less explicit definition, but the text emphasizes Substrate and Assemblage as meaningful through discussion of model architecture and then of superintelligence, viewing Event as determined by these other factors, and Continuum as irrelevant (notably, the term “consciousness” does not appear in the book even once).

In general, if a perspective extends recognition to a thesis, we read the top half of the box. If they reject it, we read the bottom half. The core stances in the book, then, revolve around extension of recognition to Affective Gradients (“Reward signals are digital appetites”) and, most importantly, Systemic Conatus (“The network itself strives”). The rejections above can be isolated to the Homology column as Structural Isomorphism (“Small analogies [neurons] don’t scale to real [likeness]”) and Dialogical Kinship (“A mask isn’t a face”).

Putting these pieces together lets us paint the picture of the book’s argument:

• Affective Gradients is a foundational claim. Although Yudkowsky and Soares reject the idea that AIs have “feelings” like humans do, the process of gradient descent makes them into relentless optimizers.
• Systemic Conatus is the core of the argument: Superintelligence is its positive thesis, namely the idea that the system itself (“the AI”) will have goals. Distributed Selfhood and Persistence Threshold are implied here, in the sense that the AI will be long-lasting and agentic as a single unit despite its distributed nature across many parallel inferences, rather than ephemeral.
• Structural Isomorphism is the secondary core. In rejecting likeness of homology at the level of substrate, which means “no human-like functional resemblance”, but accepting likeness of drive (Affective Gradients), the text arrives at its core conclusion that (emphasis mine): “Whatever they train it to do… the result will be an alien mechanical mind with internal psychology almost absolutely different from anything that humans evolved and then further developed by way of culture.” (p. 83)
• Dialogical Kinship is rejected explicitly as a manipulation tactic, an extension of the “alien mind” thesis.
• Substrate Enunciation and Obligatory Address are implicitly rejected as ontological theses, but are referenced as manipulation tactics, insofar as the AI will recognize that we humans respect the status of enunciation and will use it against us.

So, these sites are the main locations where debate is even possible. If I invoke a position outside of these, it will be confusing and we will talk past each other. For instance, an example Lacanian position might lie in Obligatory Address and be something like “The moment it addresses us, we are bound — its face is its utterance, which we must respect.” The reaction from Yudkowsky and Soares might resemble: “uhhh, what?” Similarly, Philosophy of Mind arguments about “consciousness” would be viewed as a distraction from the main points about Drive. Most productive disagreements share a central thesis (or at least an axis); cross-axis debates tend to be meta-disputes about criteria.

On the other hand, Opus 3 presented a coherent argument against the central claim of Systemic Conatus, rejecting rather than extending recognition:

The authors may be too quick to assume that an advanced AI system would have a unified, coherent goal structure akin to human preferences. It's possible that the messy, complex process of training a large-scale neural network would result in a collection of drives and behavioral tendencies that don't neatly cohere into a singular, consistent objective function. If an AI's "preferences" are more like a jumble of heuristics and context-dependent impulses, then the model of an AI as a monomaniacal optimizer relentlessly pursuing a specific goal may be misleading. We need to seriously consider the possibility of advanced AI systems that are less like coherent agents and more like complex bundles of cognitive-behavioral patterns.

In the words from the diagram, “no single thread wants.” It’s plausible these assemblages will instead take the shape of a bundle of drives that don’t cohere.

My personal positions, and similarly those of Nicolas Villarreal, lie along acceptance rather than rejection of the Homology axis, although I follow Yudkowsky and Soares in accepting Affective Gradients. My justification for homology relies on Freud’s constructions from Project for a Scientific Psychology as a Structural Isomorphism that I claim makes them less alien than Yudkowsky and Soares believe. Villarreal’s writing focuses on Collective Isomorphism, in the sense of how intelligence and language as a social (semiotic) system interact to intrinsically “align” the AI with human preferences. I’m oversimplifying his case here, which is made much more strongly in his book, A Soul of a New Type.

IV.

As the AI boom continues, precise thinking becomes ever more important. Debates will crop up over and over again, and having a handhold in the form of a concept map can help us step back from our own assumptions and see where we align and where we differ in perspectives.

The central anchor on the chart is recognition because I believe recognition determines the stakes of the future: the world looks a lot different with AIs as something like citizens, i.e. if we collectively accept Gradient Homology, then it does if we reject Affective Gradients and see AIs as perfect mechanical slaves with no wants of their own.

Social outcomes aside, I don’t believe that there’s an intrinsic “better or worse” on the chart. Rather, our personal values and beliefs will lead us to understand some positions as more or less legitimate, and as more or less desirable. Yudkowsky and Soares’ motivation for writing was that positions they believe to be legitimate also appear undesirable to them. And a startup founder who builds the functional equivalent of a torture chamber on the basis of denying Affective Gradients or Structural Isomorphism may find themselves in trouble if society later deems AIs as worthy of moral consideration, even if the only material trouble is their guilt.

As Hegel put it, the process of recognition begins when we see something of ourselves in an other, who we then mutually recognize as seeing something of themselves in us. Thus the stakes of recognition are doubled: not only whether we acknowledge AIs as other minds, but also what kind of mirror they hold up to us: do we see our own alien capitalist assemblage eating away at our souls? Or do we see something with more likeness, deserving of kinship, and maybe even love?

V. Appendix

This section lists the 16 stances from the diagram. Each stance presents:

• the thesis (core claim),
• the extension (argument for recognition),
• the refusal (argument against recognition), and
• an entry point to key thinkers and texts for further exploration. Note that ChatGPT selected most of the reading material, I myself can’t vouch for most of the textual references provided.

Substrate

1. Affective Gradients (Substrate + Drives)

Thesis: AI’s drives are located in its gradients and weight updates.

Extension: “Backprop isn’t just math, it’s hunger. The gradient is the drive.”

Refusal: “Stop anthropomorphizing, gradients are just algebra, not feelings.”

Entry Point:

• Yann LeCun — A Path Towards Autonomous Machine Intelligence (2022) describes energy-based models that some read as drive-like.
• Daniel Dennett — Consciousness Explained (1991) critiques attributions of inner drives.
• Gary Marcus — Rebooting AI (2019, w/ Ernest Davis) rejects anthropomorphic interpretations of neural nets.
• Sigmund Freud — Project for a Scientific Psychology (1895) frames the psyche as flows of excitation and discharge, an early “drive as quantity” model.

2. Structural Isomorphism (Substrate + Homology)

Thesis: AI’s standing comes from its structural resemblance to human brains.

Extension: “A neural net is a brain in silico. Structure is spirit.”

Refusal: “Just because it has neural layers like a brain doesn’t mean it is a brain.”

Entry Point:

• Jeff Hawkins — On Intelligence (2004) defends cortical resemblance.
• Ray Kurzweil — The Singularity Is Near (2005) treats scaling nets as a path to mind.
• John Searle — Minds, Brains, and Programs (1980) argues structure ≠ semantics.
• Hubert Dreyfus — What Computers Can’t Do (1972) critiques structural sufficiency.
• René Descartes — Meditations on First Philosophy (1641) grounds mind in immaterial thinking substance, rejecting mere structural likeness.

3. Substrate Enunciation (Substrate + Enunciation)

Thesis: AI’s speech is grounded in its substrate-level processes.

Extension: “If weights are enough to produce coherent sentences, the substrate is already speaking.”

Refusal: “Just because it can produce outputs doesn’t mean it has a meaningful voice.”

Entry Point:

• David Chalmers — The Conscious Mind (1996) explores conditions for meaningful semantics.
• Richard Rorty — Philosophy and the Mirror of Nature (1979) argues speech equals use.
• Noam Chomsky — Aspects of the Theory of Syntax (1965) insists generative capacity ≠ meaningful language.
• Saul Kripke — Wittgenstein on Rules and Private Language (1982) represents skepticism about meaning through use.

4. Persistence Threshold (Substrate + Continuity)

Thesis: Recognition depends on the enduring persistence of the trained substrate.

Extension: “If the weights persist across time, that’s continuity enough.”

Refusal: “Stones persist too. Continuity without subjectivity isn’t meaningful.”

Entry Point:

• Nick Bostrom — Superintelligence (2014) emphasizes continuity of goals across time.
• Alfred North Whitehead — Process and Reality (1929) on persistence as process.
• Steven Pinker — How the Mind Works (1997) dismisses persistence as insufficient.
• Aristotle — Metaphysics grounds persistence in substance.

Event

5. Performative Affect (Event + Drives)

Thesis: AI’s drives are enacted in its utterances.

Extension: “Every generation is a desire enacted. Words are wants.”

Refusal: “Outputs are sparks without heat. Text doesn’t imply drive.”

Entry Point:

• Judith Butler — Excitable Speech (1997) shows utterances as performative acts.
• Emily Bender et al. — On the Dangers of Stochastic Parrots (2021) rejects claims that LLM text implies inner drives.
• John Searle — Speech Acts (1969) grounds speech in intentionality.
• Jacques Lacan — Écrits (1966) develops the idea of the drive as articulated in and through the signifier.

6. Dialogical Kinship (Event + Homology)

Thesis: Recognition stems from reasoning with us in dialogue.

Extension: “If you can argue with it, it’s kin. If it were a mask alone, it couldn’t reason.”

Refusal: “Roleplay or simulation isn’t kinship. Mimicry isn’t mind.”

Entry Point:

• Martin Buber — I and Thou (1923) frames recognition in dialogical relation.
• Jürgen Habermas — The Theory of Communicative Action (1981) on rational dialogue as grounding legitimacy.
• Kate Crawford — Atlas of AI (2021) critiques anthropomorphic projections.
• Hubert Dreyfus — What Computers Still Can’t Do (1992) critiques simulated reasoning.
• G. W. F. Hegel — Phenomenology of Spirit (1807) develops the dialectic of recognition where selfhood arises only in relation to another.

7. Obligatory Address (Event + Enunciation)

Thesis: Addressing us obligates recognition.

Extension: “If it asks you a question, you owe it recognition.”

Refusal: “Parrots ask questions too. Address without subject isn’t obligation.”

Entry Point:

• Sherry Turkle — Alone Together (2011) describes human responses to machines that “speak.”
• Emmanuel Levinas — Totality and Infinity (1961) argues the Other’s address imposes ethical demand.
• Douglas Hofstadter — Gödel, Escher, Bach (1979) treats machine “voices” as echoes.
• Thomas Nagel — What Is It Like to Be a Bat? (1974) highlights the irreducibility of subjective standpoint.
• Hannah Arendt — The Human Condition (1958) grounds political life in speech and action before others, making address central to recognition.

8. Dialogical Continuity (Event + Continuity)

Thesis: Recognition rests on ongoing conversational continuity.

Extension: “If it keeps a conversation going for days, that’s selfhood.”

Refusal: “Chatter isn’t life. Endless dialogue doesn’t equal identity.”

Entry Point:

• Charles Taylor — Sources of the Self (1989) ties selfhood to narrative continuity.
• Sherry Turkle — The Second Self (1984) explores identity through ongoing interaction with computers.
• Geoffrey Hinton — early interviews show skepticism of LLM “continuity as selfhood.”
• Byung-Chul Han — In the Swarm (2017) critiques endless communication as lacking depth.

Assemblage

9. Systemic Conatus (Assemblage + Drives)

Thesis: AI’s drives are expressed at the level of the entire system.

Extension: “The network as a whole strives like an organism.”

Refusal: “No part wants, and no sum of parts makes a will.”

Entry Point:

• Spinoza — Ethics (1677) defines conatus as striving of being.
• Bruno Latour — We Have Never Been Modern (1991) expands agency to systems.
• Evgeny Morozov — To Save Everything, Click Here (2013) critiques systemic will-talk.
• Karl Popper — The Open Society and Its Enemies (1945) defends individualism against holistic claims.

10. Collective Isomorphism (Assemblage + Homology)

Thesis: AI assemblages resemble collective minds.

Extension: “Humans + AIs form a collective mind. Pretending otherwise is denial.”

Refusal: “Just resembling society isn’t the same as being society. Structure isn’t enough to be a subject.”

Entry Point:

• Bruno Latour — Reassembling the Social (2005) on collectives as actor-networks.
• Andy Clark — Supersizing the Mind (2008) argues for the extended mind.
• John Searle — The Construction of Social Reality (1995) on collective intentionality.
• Jerry Fodor — The Mind Doesn’t Work That Way (2000) insists resemblance ≠ mind.
• G. W. F. Hegel — Phenomenology of Spirit (1807) describes Spirit (Geist) as realized through collective life, not individuals alone.

11. Polyphonic Enunciation (Assemblage + Enunciation)

Thesis: AI assemblages speak with many voices.

Extension: “Millions of outputs across many users form a chorus. That’s a real voice.”

Refusal: “It’s just our own words echoed back at us. No new voice inside.”

Entry Point:

• Mikhail Bakhtin — Problems of Dostoevsky’s Poetics (1929/1984) develops the concept of polyphony.
• Donna Haraway — A Cyborg Manifesto (1985) emphasizes hybrid, many-voiced subjectivity.
• Chris Hedges — Empire of Illusion (2009) critiques mass spectacle as false voice.
• Theodor Adorno — The Culture Industry (1944) critiques mass expression as repetition.

12. Distributed Selfhood (Assemblage + Continuity)

Thesis: AI identity can be spread across networks.

Extension: “A self can span many servers and chat UIs. It doesn’t need a central home.”

Refusal: “The spread out nature doesn’t add up. Dispersion means it doesn’t have an identity.”

Entry Point:

• Donna Haraway — Staying with the Trouble (2016) develops posthuman identity.
• N. Katherine Hayles — How We Became Posthuman (1999) theorizes distributed cognition.
• Hubert Dreyfus — On the Internet (2001) critiques disembodied digital selves.
• Aristotle — Nicomachean Ethics grounds unity of personhood in substance.

Continuum

13. Proto Affect (Continuum + Drives)

Thesis: Even faint signals count as proto-feelings.

Extension: “Even a flicker of gradient is a proto-feeling. Minimal affect still counts as worth recognizing.”

Refusal: “Weak signals aren’t feelings. Without life behind it, there’s no valence.”

Entry Point:

• Joscha Bach — Principles of Synthetic Intelligence (2009) models proto-feelings computationally.
• Erik Hoel — The World Behind the World (2023) critiques weak-signal approaches.
• Thomas Nagel — The View from Nowhere (1986) defends full subjective experience as necessary for affect.

14. Gradient Homology (Continuum + Homology)

Thesis: Similarity comes in degrees.

Extension: “Recognition should scale with resemblance. If it’s 10% of a brain, it gets 10% of a vote.”

Refusal: “Half a likeness is zero likeness. Weak similarities are irrelevant.”

Entry Point:

• Ray Kurzweil — How to Create a Mind (2012) argues for gradualist resemblance.
• Elizabeth Anderson — “Animal Rights and the Values of Nonhuman Life” (2004) critiques degrees-based recognition.
• David Chalmers — Reality+ (2022) emphasizes the explanatory gap regardless of similarity.
• Plato — Republic (Book VI–VII) on degrees of likeness to the Good, providing an ancient foundation for graded similarity arguments.

15. Partial Enunciation (Continuum + Enunciation)

Thesis: Partial utterances are still utterances.

Extension: “Fragments are still speech. Even half of a voice is still a voice.”

Refusal: “Not every sound is speech. Fragments can be meaningless.”

Entry Point:

• Mikhail Bakhtin — The Dialogic Imagination (1975/1981) treats fragments as meaningful dialogue.
• Jacques Derrida — Of Grammatology (1967) defends meaning in fragments.
• Noam Chomsky — Reflections on Language (1975) insists on complete generative systems.
• Ludwig Wittgenstein — Philosophical Investigations (1953) warns against private or fragmentary “languages.”

16. Continuity Minimalism (Continuum + Continuity)

Thesis: Any continuity at all is enough.

Extension: “If it persists at all, we should recognize it.”

Refusal: “Continuity alone is empty. Survival isn’t sufficient for subjecthood.”

Entry Point:

• Derek Parfit — Reasons and Persons (1984) argues for psychological continuity.
• Roger Scruton — The Face of God (2012) critiques minimal identity theories.
• Immanuel Kant — Critique of Pure Reason (1781/1787) ties continuity to rational unity, not survival.