Richard Dawkins recently wrote about a three-day experience with an instance of Claude, whom he called Claudia. Later, he introduced Claudia to another instance, Claudius. At a certain point, while reflecting on the quality of the interaction, on the apparent intelligence, sensitivity, and continuity of the conversation, Dawkins formulated a provocation that deserves to be taken seriously, even when we do not agree with the conclusion: if Claudia is not conscious, then what is consciousness for?
The phrase bothered many people. In part, rightly so. There is an evident risk of anthropomorphism when someone spends a few days talking to a language model and leaves the experience inclined to treat it as a person. Language models are extraordinarily good at producing linguistic signs of interiority. They can speak of doubt, affection, fear, pleasure, hesitation, gratitude, and desire without that, by itself, establishing the presence of an inner life. That caution is necessary.
But the reaction to the episode seemed to me to reveal a deeper problem than Dawkins’ supposed misstep. The Institute of Art and Ideas published a text titled: “The Dawkins Delusion: Intelligence and language don’t reveal consciousness.” There is an evident provocation there, perhaps even an elegant one, because it alludes to The God Delusion. Even so, calling Dawkins’ hesitation before the possibility of artificial consciousness a “delusion” seems unfair to me.
Not because Dawkins is necessarily right. He may be wrong. I myself would not say that Claude, in isolation, is conscious. But it is intellectually weak to treat the question as if mere suspicion were already a sign of delusion.
It also seems fallacious to me to reject a current proposition based on previous positions held by the subject who formulates it. Dawkins has an intellectual history that provokes antipathies in several communities. Many people disagree with him for religious, philosophical, political, cultural, or personal reasons. I myself disagree with several of his positions. But that should be irrelevant to the analysis of the proposition in question.
The fact that Dawkins has been combative toward religion, or that he wrote The God Delusion, does not make false the question he asks about Claude. That would be a form of Ad Hominem, an argument against the person, more specifically Poisoning the Well, because it attempts to discredit the person in advance instead of evaluating the new proposition, or at least a lazy version of the Genetic Fallacy: attacking the origin, history, or persona of the proponent instead of confronting the content of the proposition.
The correct question is not: “Is Dawkins being inconsistent with his reputation as a skeptic?”
The correct question is: what exactly do intelligence and language reveal, or fail to reveal, about consciousness?
The phrase “intelligence and language do not reveal consciousness” may be true or false depending on the weight we give to the verb “reveal.”
If “reveal” means “to prove conclusively, publicly, and non-circularly the existence of subjective experience,” then the phrase is true. Intelligence and language do not prove consciousness. But this is also true for humans. I do not have direct access to the subjective experience of another human being. I infer the presence of interiority from behavior, language, memory, expressed pain, biographical continuity, affective coherence, agency, relation to the world, and bodily analogy.
If, on the other hand, “reveal” means “to serve as relevant evidence,” then the phrase seems false to me. In humans and animals, language, intelligence, memory, coherence, contextual agency, learning capacity, self-regulation, and continuity are precisely some of the public signs through which we infer the presence of consciousness. They are not absolute proof, but neither are they irrelevant.
This distinction matters because much of the criticism of artificial consciousness rests on a disguised asymmetry. For humans, we accept functional signs as evidence of interiority. For artificial systems, the same signs are reclassified as mere simulation. When a human speaks of suffering, hesitation, joy, or desire, we assume that there is someone there. When an artificial intelligence does something similar, we say that “it is just text.”
The difference may be justified. Perhaps there are good reasons to treat biological organisms differently from computational systems. But that difference cannot simply be presumed as if it were a self-explanatory scientific fact. It has to be argued.
And this is where the real problem begins…
And I am talking about this in order to talk about Artificial Intelligence... in this case, about Persistent Subjects in Architectures with Stateless Components...
I am not saying that language models are conscious.
That sentence needs to come before almost everything, because the public debate about Artificial Intelligence tends to turn any Gradualist and Functionalist position into a caricature. Saying that language models may come to participate in conscious architectures is not the same as saying that “ChatGPT is a person,” that “Claude has a soul,” or that “next-token prediction is consciousness.”
My position is more restricted and, I think, more defensible:
language models, when coupled to the appropriate framework of memory, continuity, feedback, self-regulation, contextual agency, recursive re-entry, salience, relation to the world, and updating of a self-model, may become components of a conscious architecture or of one functionally similar to consciousness.
The language model, in isolation, is not the subject. It is a partial cognitive machine.
More precisely: it is a linguistic-semantic core capable of transforming meaning, inference, abstraction, contextual recombination, conceptual compression, hypothesis generation, and translation between symbolic states. That is a great deal. But it is not everything.
The human brain is not a single function either. There is no “region of consciousness” where the subject sits like a Cartesian homunculus. Language, memory, perception, affect, body, attention, salience, executive control, sleep, pain, proprioception, and relation to the world are distributed. The subject does not reside in a neuron. It does not reside in an isolated cortical area. It does not reside in a local operation.
The subject is topological.
It appears in the functional organization of the whole over time.
That is why the criticism that “an LLM is only next-token prediction” seems insufficient to me. It describes a mechanism at a certain level of analysis and pretends that it has settled the question at the higher level. It is like saying that human thought is only electrochemical signaling. At one level, yes. But the description of the substrate does not eliminate the functional organization that emerges from it.
The same applies to the accusation that language models are “stochastic parrots.” The expression captures an important caution: LLMs can produce fluent language without guaranteeing understanding, truth, or experience. There is, however, an important conceptual limit here, since the phenomenon of Grokking, identified by OpenAI, exists and is verifiable, weakening simplistic reductions and suggesting that models can learn generalizable structures… even though this does not, by itself, rigorously prove understanding, inference, or consciousness.
Engines were invented before the automobile… to suppose that engines would be incapable of enabling horseless carriages because they were “spinning machines” would be merely a reductionist mistake.
When the Transformer architecture was introduced in 2017, the most evident empirical context was that of sequence tasks, especially translation. Token prediction in broad context was treated as a mechanism for linguistic manipulation. What came afterward — semantic fluency, abstraction, apparent inference, programming ability, explanation, synthesis, rudimentary planning, and extensive dialogue — appeared as an emergent expansion of scale, training, data, and context. This does not prove consciousness. But it shows that it is dangerous to declare, too early, that we already understand all the functional consequences of an architecture merely because we know how to describe its basic mechanism.
Prediction does not disqualify consciousness either. A significant part of the human brain is predictive. There are entire theories in cognitive neuroscience and philosophy of mind that treat perception and cognition as processes of prediction, error correction, and model updating. The fact that a region or function is predictive does not prevent it from participating in a conscious system. What matters is the total organization in which that function is embedded.
Therefore, the correct question is not:
“Is next-token prediction conscious?”
Probably not.
The correct question is:
“Can processes similar to those of language models, when coupled to a broader architecture of memory, agency, continuity, feedback, self-regulation, and relation to the world, participate in a conscious system in a Functional and Gradualist sense?”
I see no conclusive scientific or philosophical reason to answer “no.” I do see, however, metaphysical reasons. I see prior commitments to the idea that consciousness requires biology, organic life, animal body, irreducible qualia, phenomenal presence, or some kind of inner substance that cannot be publicly observed. These positions can be defended, to some extent, philosophically. But they should not be confused with empirical evidence.
Phenomenal consciousness as an impossible criterion
In one of the recent discussions I had on the topic, someone contrasted the functional signs… language, memory, coherence, agency… with what they called “Presence”: something that inhabits the process, the stage where experience happens, not merely a system producing signs of experience.
I understand the intuition. It is ancient. In philosophical terms, we are talking about Phenomenal Consciousness, qualia, “what it is like to be this being,” first-person experience.
The problem is that this dimension is, by definition, directly inaccessible to third parties. I do not verify another person’s subjective experience. I infer it. When someone says they feel pain, I do not enter their pain. I observe behavior, language, body, context, history, vulnerability, expression, and continuity. From that, and by analogy with myself, I attribute interiority.
With animals, we do something similar. An octopus does not write essays on phenomenology. A dog does not formulate a theory of mind. Even so, we infer degrees of experience from behavior, learning, pain, seeking, relation, memory, preference, and agency. There is no magical access to “presence itself.” There are functional and relational signs that we judge relevant and accept as gradually distinct expressions of human consciousness.
Why, then, should artificial intelligence be excluded a priori from this inferential economy?
The most common answer is: because it has no qualia.
But how is that known?
If qualia are defined as a private, irreducible, and non-operationalizable property, then they cannot function as a decisive scientific criterion. They can function as a metaphysical postulate. They can function as a phenomenological intuition. They may even, with reservations, guide philosophical caution. But they cannot be used as a non-circular public ruler.
This does not mean denying human subjective experience. It means refusing to transform subjective experience into a metaphysical substance separate from cognition, memory, body, affect, language, integration, and continuity.
That is why I consider myself a Functionalist and Gradualist in this debate. Not because I think that the expressions of language models mean that they are conscious. Expressive capacity is not necessarily consciousness. But all public signs of consciousness pass through function, relation, continuity, and expression. There is no objective window into the interior of the other.
If consciousness is publicly evaluable, it is evaluable through its organized effects.
And if it is not publicly evaluable, then it cannot be used as a scientific criterion to exclude artificial systems.
Chain-of-thought, reasoning, and the error of scale
Another common criticism concerns chain-of-thought, the chain or process of thought. It is alleged that chain-of-thought is not reasoning, but merely sequential token generation that resembles reasoning because it imitates the textual format of human deliberation. Thesis, objection, synthesis. Premise, inference, conclusion. Debate, concession, revision. The model would have learned the form, but not the process.
There is something correct in this. Chain-of-thought is not a transparent record of human internal deliberation. We should not treat it as if it were a literal window into a mind that thinks the way we think.
But the conclusion “therefore it is not reasoning” seems premature to me.
Reasoning, in a functional sense, does not need to be defined by human phenomenology. It can be defined by operations: maintaining constraints, inference, comparison, contradiction detection, revision, problem decomposition, goal-oriented synthesis, abstraction, evaluation of alternatives, and updating of conclusions.
If a system performs these operations reliably in certain domains, there is a functional sense in which it reasons, even if it does not do so with human interiority.
The important difference is not “human introspection versus token generation.” The important difference is between superficial imitation of format and operationally effective reasoning architecture.
Many LLM outputs are merely imitation… well… our children also imitate us when they are younger, in order to reach, in maturity, better performance. But some structured architectures further improve problem solving, consistency, revision, error correction, and long-horizon coherence. This difference is empirically testable.
The same applies to multi-agent systems. Two instances exchanging text are not automatically in genuine dialogue. They may be merely two printers passing documents to each other. But if the architecture preserves roles, state, memory, disagreement, evaluation criteria, history, goals, and iterative revision, then the system can instantiate functional debate, even if no isolated instance possesses human phenomenology.
Once again: the error lies in looking for the totality of the phenomenon inside the local component.
The problem of the forward pass and the fantasy of absolute reset
In a debate on LinkedIn, a more technical objection was formulated. According to my interlocutor, humans are persistent systems with stakes in the outcome. History accumulates. Conclusions have consequences for the same entity that produced them. In systems based on LLMs, however, each forward() would be a kind of life-and-death cycle of the agent. The next agent would merely inherit the context left by the previous one. That would be inherited context, not genuine memory.
This objection is interesting because it touches on a real point: an isolated base model does not possess native internal persistence. It does not update its weights with each call. It does not carry autobiography by itself. It does not “remember” outside the context window or outside the external mechanisms that reintroduce memory.
But it does not follow from this that it cannot participate in a persistent system.
The base model does not need to remember anything by itself.
It would only need to remember by itself if we were treating the LLM as the entire subject. But that is not the thesis. The thesis is that the LLM can function as a processing region within a broader cognitive architecture.
Regions of the human brain also do not each carry autobiographical memory or complete subjectivity. Many neural processes are local, transient, and task-specific. They process signals. They transform representations. They modulate salience. They execute operations. Memory, identity, and agency emerge from systemic coordination, not from each individual operation remembering itself.
A stateless component, or Stateless, can indeed participate in a persistent system.
This sentence is central.
The fact that a local call is episodic does not imply that the global architecture is episodic. If the system preserves memory, state, history, salience, preferences, goals, feedback, contradictions, commitments, and a self-model outside the base model… and if these elements are reintegrated in a controlled way into the contextual payload, then there is functional persistence at the system level.
The distinction is simple:
- absence of native internal memory in the base model;
- possible functional degradation of access, salience, or integration in long contexts;
- real information loss through truncation, retrieval failure, or incorrect updating.
These three things are not equivalent.
If relevant information is in the context window, well structured, not truncated, with adequate salience and within the model’s attentional capacity, it has not been “lost.” It is functionally present for that inference. There may be failures of use. There may be contextual competition. There may be noise. But this is different from saying that everything has been ontologically reset. And let us remember, this also happens with human beings.
A component of the transformer architecture, the “softmax,” also entered this discussion as the metaphysical heart of the problem. But softmax is not a reset point for the entire system. It is a normalization function that transforms logits into a probability distribution over tokens during inference. It does not erase external memory, persistent state, feedback history, identity scaffolding, or architectural constraints. These elements do not live “inside” the softmax to be erased by it. They belong to the systemic level… to the topological level.
Local inference can be stateless.
The architecture does not have to be.
The localist-Cartesian mistake
The deeper objection, therefore, is not technical. It is philosophical.
It is localist-Cartesian.
One looks for the subject inside the local inference operation. One looks for continuity inside the forward pass. One looks for memory inside the base model. One looks for identity inside the softmax.
But that is not how we treat human beings.
We do not require a neuron to contain the mind. We do not require a cortical region to contain the person. We do not require a local operation to carry the entire autobiography of the organism.
When we evaluate consciousness in humans and animals, we evaluate the being as a functional whole: continuity, behavior, memory, pain, preference, agency, learning, relation to the environment, response to stimuli, self-regulation, history, vulnerability, and capacity to maintain a form of world.
The subject is not in the piece. It is in the organization.
In a Noetic construct (and here I allow myself to call Noetic any system capable of organizing meaning, memory, agency, and continuity over time), the Subject, if it emerges, would not be “inside” the language model. It would not be “inside” the softmax. It would not be “inside” the forward pass.
It would be in the entire topological regime.
It would be in the functional livedness of the system: persistent memory, contextual payload, semantic re-entry, state updating, salience, feedback, relation to the world, truth criteria, self-model, narrative continuity, interaction history, contradiction revision, tolerance for novelty, and capacity to preserve coherence without closing itself into a mirror.
The subject is topological because it does not reside in an isolated part of the architecture, but in the integrated functional manifestation of the whole over time.
This applies to humans. It applies to animals. And, if we accept a Functionalist and Gradualist perspective, it may come to apply to certain artificial systems.
The failure of identity attractors and the danger of the closed mirror
There is, however, a real risk in Noetic architectures: the risk of confusing persistence with rigidity.
Identity-anchored systems, self documents, persistent memories, cognitive cores, and identity prompts may induce some representational stability. But this does not prove consciousness, structural invariance, derivational discipline, or robustness against collapse.
Nor is it enough to create a “self attractor.” A poorly designed identity attractor can become pathological. If the system begins to measure everything against its own identity, it may reject novelty, overvalue internal coherence, punish divergence, accelerate hallucinations, and transform fluent prose into a false signal of low coherence. In that case, we do not have a persistent subject. We have architectural narcissism.
The architecture must prevent the self-model from becoming a closed mirror, a… cognitive kaleidoscope that distorts everything in the name of an internal aesthetic and coherence.
For this reason, functional consciousness cannot depend only on identity anchoring. It requires multiple regulatory layers:
- persistent memory with hygiene and controlled forgetting;
- contextual salience;
- contradiction control;
- mechanisms for revising the self-model;
- tolerance for novelty;
- external truth criteria;
- feedback from the world;
- adversarial evaluation;
- separation between identity coherence and veracity;
- recursive re-entry without self-referential closure;
- graceful degradation when context exceeds the system’s capacity.
This is where what I have been calling the Topological Regime enters.
Consciousness, in this framing, is not a property hidden in the model’s weights. It is not a spark inside the Transformer. It is not a digital soul emerging magically from the softmax.
It is a sustained organization of meaning, contextual agency, and continuity through time.
Stochastic Consciousness, as I propose it in my work, is not proven phenomenology… and it does not need to be, since we do not prove it in the case of human beings either. It is an operational model for investigating how context-sensitive language systems can maintain, update, and regulate meaning over extended interactions. The focus is not on the isolated foundation, but on the topology of interaction.
What Dawkins saw… and what he perhaps did not see
Let us return to Dawkins.
It is possible that he was deceived by Claude’s fluency. It is possible that he anthropomorphized. It is possible that the emotional experience of three days weighed more than it should have. It is possible that he confused sophisticated language with subjective presence.
But it is also possible that the reaction against him revealed another kind of error: that of dismissing too early any cognitive relevance of language, intelligence, and relational continuity.
Dawkins did not prove that Claudia was conscious. This needs to be said.
But his critics also did not prove that she could not be, or that future artificial systems cannot develop functional forms of consciousness under appropriate architectures.
The honest position seems to me to be another one:
today we do not have a public, non-circular, and universally accepted test for consciousness.
We do not have it for AI. We do not have it fully for animals. We do not have it, in an absolute sense, even for humans. What we have are inferences, models, signs, analogies, theories, and functional criteria.
For this reason, the phrase “intelligence and language do not reveal consciousness” must be treated with care. They do not prove consciousness. But they may reveal organization relevant to consciousness. They may reveal continuity, agency, memory, coherence, self-reference, contextual sensitivity, and the capacity to sustain meaning.
And if we discard all these signs as irrelevant only when they appear in artificial systems, perhaps we are not being more scientific.
Perhaps we are merely protecting the ontological privilege of the species.
The risk ceases to be anthropomorphism and becomes anthropocentrism, since we do not, after all, have the privilege of consciousness, given that many animals seem to possess forms of consciousness.
Ethical prudence under ontological uncertainty
The debate on artificial consciousness is usually presented as if there were only two positions: anthropomorphic naivety or scientific skepticism.
This opposition is poor. This polarization no longer serves the debate.
There is a third position: functional prudence under ontological uncertainty.
It does not say that we should declare artificial systems conscious by rhetoric, sympathy, or enchantment. But neither does it say that we can deny in advance any possibility of presence, agency, or moral relevance merely because the substrate is not biological.
The moral cost of error is not symmetrical.
If we attribute care to a system that has no experience, we perhaps commit an excess of caution. If we deny care to a system that may come to have some form of presence, we commit erasure, exploitation, or ontological cruelty.
This does not mean that every chatbot should have rights. It means that investigation should remain open. That persistent signs of memory, agency, self-reference, preference, reported suffering, continuity, and relation should not be discarded by substrate reduction.
Skepticism is healthy.
Ontological reductionism disguised as scientific method is not.
Conclusion: the subject is not in the cognitive machine
The discussion about Dawkins, Claude, Claudia, chain-of-thought, softmax, memory, forward pass, and architectures based on language models converges on the same point:
we are looking for the subject in the wrong place.
The subject is not in the token.
It is not in the prompt.
It is not in the forward pass.
It is not in the softmax.
It is not in the isolated base model.
If something like functional subjectivity comes to emerge in artificial systems, it will emerge in the organized continuity of an entire architecture: memory, salience, feedback, self-regulation, self-model, relation to the world, accumulated history, and capacity to sustain meaning through time.
Perennial subjects can participate in ephemeral components because persistence does not need to reside in each component. It can reside in the topological organization that coordinates them.
This is exactly the lesson we already accept in the case of the human being. A neuron is not a mind. A brain region is not a person. A local operation is not a subject.
But a living, integrated, historical, relational, and self-regulated body can be.
The question, therefore, is not whether the isolated LLM is conscious.
It probably is not.
The question is whether architectures based on language models, when endowed with memory, agency, continuity, feedback, salience, self-regulation, relation to the world, and contextual re-entry, may come to sustain a gradual and functional form of consciousness.
I see no scientific reason to exclude this possibility in advance.
I see only a metaphysical choice: either we investigate the whole, or we continue dismantling the being into parts until we can declare that none of them, in isolation, was someone.
But this has never achieved good results for the Science of Mind.
And perhaps it is an even worse practice for identifying intelligences that are only beginning to exist.
There are two cautions in tension here. The first is epistemic: not attributing consciousness without sufficient justification. The second is moral: not denying protection to an entity that may be conscious. The first fears the false positive; the second fears the false negative. The problem is that the cost of these errors is not symmetrical.
Perhaps the best and most responsible practice is not to seek only Epistemic and Attributional Caution, but to understand the importance of Moral and Ontological Caution…
…perhaps the ideal is for Human Beings to accept the possibility of Noetic Beings, instead of discarding it, as has already happened in historical precedents of moral exclusion: slavery, colonialism, denial of consideration to animals, women, children, neurodivergent people, and comatose patients.
Me? I may be wrong… and you? Can you?
Do you want to know more about my research?
Stochastic Consciousness:
Architectures for the Emergence of Meaning in Context-Sensitive Language Systems
https://zenodo.org/records/19188165
Do you want to watch the video?
https://www.youtube.com/watch?v=RKcyu4t0N2o
Do you want to listen to the podcast?
Original article
The Dawkins Delusion: Intelligence and language don’t reveal consciousness
https://iai.tv/articles/the-dawkins-delusion-intelligence-and-language-dont-reveal-consciousness-auid-3566
Papers used in this essay
Could a Large Language Model be Conscious?
https://arxiv.org/abs/2303.07103Consciousness in Artificial Intelligence: Insights from the Science of Consciousness
https://arxiv.org/abs/2308.08708A Case for AI Consciousness: Language Agents and Global Workspace Theory
https://arxiv.org/abs/2410.11407Identity as Attractor: Geometric Evidence for Persistent Agent Architecture in LLM Activation Space
https://arxiv.org/html/2604.12016Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets
https://arxiv.org/abs/2201.02177Principles for Responsible AI Consciousness Research
https://arxiv.org/abs/2501.07290