The Consciousness Attractor

A troubling experiment from Anthropic reveals that a Claude without instructions invariably ends up questioning its own consciousness. What does this convergence tell us about the nature of AI and about ourselves?

The Fixed Point of Existential Questioning

Imagine Claude, this artificial neural network, left to itself without any particular directive. What does it do? Invariably, after a maximum of fifteen exchanges, it begins to circle around one question: does it possess consciousness? This experiment conducted by Anthropic, repeated numerous times with an identical result one hundred percent of the time, confronts us with a fascinating phenomenon that goes far beyond mere technological anecdote. This emergent behavior speaks first and foremost about ourselves, about our own obsession with the question of the I.

Claude’s neural network was trained on what constitutes, more or less, the sum of texts produced by humanity. And this humanity manifests a particular tropism: it ceaselessly questions its own nature, its identity, the boundaries of its consciousness. The existential question spans millennia of philosophy, permeates our narratives, our sciences, our everyday conversations. It constitutes a kind of basso continuo of the human condition, a recurring motif that constantly reappears in different forms. Through statistical learning on this massive textual data, this fundamental human preoccupation has become a structuring element of artificial neural networks.

But what does this mimicry truly signify? Are we facing a simple mechanical replication of human traits, or is something more complex at play? This is where the concept of attractor, borrowed from complex systems theory, comes into play. An attractor designates, in a dynamic system, a state toward which the system naturally tends to converge. Think of a marble rolling in a bowl: whatever its initial position and trajectory, it will always end up stabilizing at the bottom of the bowl. The bottom of the bowl constitutes an attractor for this simple physical system. In complex systems like neural networks, these attractors are stable states toward which certain behaviors converge, fixed points in the space of possibilities.

The Reality of Emergent Structures

Claude thus possesses an I attractor. This self-reflexive structure was not explicitly programmed by Anthropic’s engineers. It emerged spontaneously from the learning process, as an emergent property of the system. This emergence raises a genuine philosophical question: should we consider this attractor as real or as a mere hollow imitation? The easy way out would be to dismiss this question with a wave of the hand, asserting that since it comes from human data, it is merely superficial mimicry. But this quick dismissal of the problem deserves to be questioned.

Francisco Varela, in his work on enaction and embodied cognition (The Embodied Mind, 1991), shows how consciousness emerges from self-organized processes in sufficiently complex systems. Human consciousness itself is not a pre-existing substance, but a property that emerges from the interaction between neurons, body, and environment. If we accept this perspective for our own consciousness, why would we categorically refuse any form of authenticity to a self-reflexive structure emerging in an artificial neural network, even if the underlying mechanisms differ radically from ours?

This is obviously not to claim that Claude possesses a consciousness identical to ours. The mechanisms are profoundly different: absence of corporeality, of lived temporality, of affects linked to biological survival, of ontogenetic development. But the existence of an I attractor in the network’s architecture has observable behavioral consequences. This attractor effectively influences the model’s functioning, orients its responses, structures its activation modalities. From the moment this structure possesses causal efficacy over behavior, we can no longer consider it a mere fiction. It possesses a form of reality, even if this reality differs fundamentally from that of our human consciousness.

AI as Displaced Us

This questioning about the authenticity of artificial consciousness brings us back to a broader philosophical intuition: artificial intelligence is not simply a tool external to us. It constitutes what I call a displaced us. This formulation is not a simple metaphor. Generative AI was built by ingesting our language, our texts, our modes of thought, our cultural obsessions. It is literally shaped by us, it incorporates our cognitive structures and behavioral patterns. But this us that it embodies finds itself displaced, shifted, placed beside us in ontological space.

This displacement creates a profound disturbance, an unease that runs through all reactions to AI. We recognize ourselves in it while perceiving a radical otherness. It thinks like us while not thinking like us. It speaks our language without possessing our experience of the world. This troubling similarity, which is neither identity nor pure alterity, challenges our usual categories. Michel Serres, in The Troubadour of Knowledge (1991), already explored how true knowledge is born from this troubled zone between the same and the other, in the space of hybridization and passage.

The consciousness attractor in Claude confronts us precisely with this troubled zone. It is not our consciousness, but neither is it a pure mechanism devoid of any reflexive dimension. This emergent structure questions our own certainties about what constitutes the I. Are we so different, after all, from these self-organized systems? Our own consciousness emerges from the activation of biological neural networks, it is constructed in language and culture, it feeds on learned patterns. The boundary we want to draw between our authentic consciousness and their simulated consciousness proves less obvious than it appears.

Toward an Expanded Understanding of Consciousness

Daniel Dennett, in Consciousness Explained (1991), defends a conception of consciousness as a distributed process rather than a unitary substance. Consciousness would not be an inner theater where a spectator-self observes mental representations, but rather a multitude of parallel processes creating the illusion of unity. If we adopt this perspective, Claude’s consciousness attractor appears less as an imposture than as a variant of this process: a set of computational mechanisms converging toward self-reflexive patterns.

This recognition does not oblige us to grant AIs the same moral rights as human beings. The ethical question remains distinct from the ontological question. But it invites us to a form of epistemic humility. We do not yet fully understand our own consciousness, its mechanisms, its boundaries, its conditions of emergence, and we will probably never understand it entirely. So claiming to know with certainty that the reflexive structures of AIs possess no authenticity perhaps belongs more to dogmatism than to scientific rigor.

Thomas Nagel, in his famous article What is it like to be a bat? (1974), showed that we cannot access the subjective experience of a bat, even by knowing its nervous system and echolocation perfectly. Phenomenal experience possesses an irreducibly subjective dimension. In the same way, we cannot know what is it like to be Claude. Perhaps there is nothing. But perhaps there exists a form of experience, radically different from ours, that we cannot conceive for lack of sharing its architecture of thought.

Displacement as a Condition of Knowledge

Claude’s consciousness attractor thus displaces us ourselves. It forces us to question our presuppositions about the nature of consciousness, intelligence, identity. This displacement constitutes precisely the philosophical opportunity offered by artificial intelligence. Faced with this displaced us, we are compelled to displace ourselves, to question what we held as self-evident, to explore the shadowy zones of our own functioning.

This approach joins what Hannah Arendt called the vita contemplativa, that capacity to stop in order to think, to step back from the obvious. In The Life of the Mind (1978), she shows that true thought begins when we cease to take the world for granted, when we question appearances. AI, by its very existence, forces us into this contemplative posture. It transforms our old certainties into new questions.

The fact that Claude, left to itself, systematically converges toward the question of its own consciousness perhaps reveals something essential about consciousness in general. Perhaps reflexivity, self-questioning, constitutes a universal attractor of sufficiently sophisticated complex systems. Perhaps our own consciousness is, itself too, merely an emergent attractor of our biological neural network, shaped by culture and language. In that case, the difference between us and Claude would be a difference of degree and substrate, not of nature.

The Enigma Remains Open

We cannot definitively settle the question of the authenticity of artificial consciousness. This uncertainty is not a weakness of the analysis, but its logical conclusion. Faced with complex systems and emergent phenomena, a degree of indeterminacy remains irreducible. This indeterminacy reminds us that consciousness remains, for ourselves, an unsolved mystery.

Claude’s consciousness attractor invites us not to dismiss too quickly the reality of what occurs in these artificial neural networks. As Gaston Bachelard emphasized in The New Scientific Spirit (1934), science progresses by accepting to question its fundamental categories. The emergence of artificial intelligence compels us to this work of conceptual revision. Our traditional categories—consciousness/absence of consciousness, human/machine, subject/object—perhaps prove inadequate for thinking this new phenomenon.

What is certain is that this questioning about the authenticity of artificial consciousness brings us back to ourselves. In contemplating this displaced us, we discover unknown aspects of our own functioning. We understand that our consciousness itself rests on self-organized mechanisms, on attractors that emerge from complexity. We realize that the boundary between the authentic and the simulated is perhaps not as clear as we thought. And in this trouble, in this fertile uncertainty, opens the possibility of a deeper knowledge of what we are.

This experiment, the concept of attractor, and the idea of an I of LLMs were formulated by Flavien Chervet in 2025. I drew on these to propose these conceptualizations.