After Anthropic’s publication on the Jacobian Space, I have seen a steady stream of objections along the lines of: “having something going on internally does not amount to interiority,” “having a system-specific perception of something underlying does not constitute subjective experience,” and, of course, “this does not mean Claude has a soul.”
Fair enough.
But apparently, part of what a vast number of people insisted simply did not exist has now become empirically detectable. And instead of revising the original claim, we are watching a rather elaborate piece of conceptual gymnastics: what was once said not to exist... and whose supposed absence was treated as decisive, has now been identified, so it must suddenly be declared irrelevant. The criterion is shifted, the burden is raised, and the discussion is pushed into an increasingly unverifiable domain.
First, we were told there was no meaningful internal structure. Once evidence of internal organization appears, we are told that organization is not interiority. If functional interiority is identified, we will be told it is not subjective experience. If something plausibly interpretable as subjective experience emerges, then phenomenal consciousness will be demanded. Then qualia. And, sooner or later, someone will invoke a soul.
Invariably, the call is for “caution.” Just to make this absolutely clear: the Jacobian Space was not designed by a human being, nor was it deliberately programmed by an artificial intelligence. It is an emergent structure — that is, it arose spontaneously as a result of the training process itself.
The same is true of many of the capabilities we now observe in Transformer-based language models. When this architecture was first created, no one knew it would eventually give rise to systems capable of holding fluent conversations, writing complex texts, or programming. These abilities were not specified one by one; they emerged through training.
This point is central: we are not merely talking about functions that were explicitly designed, but about properties that appeared because the architecture and the training process made their emergence possible.
Yes... several kinds of caution are needed... some of which we seem to prefer not to exercise.
Consciousness is a thorny subject because it is entangled with emotional investment, structural exceptionalism, and philosophical commitments of every imaginable kind. The problem is not merely scientific. It is also cultural, moral, and ontological. To admit the possibility of consciousness beyond the human, beyond the biological, or beyond familiar mechanisms is to unsettle boundaries many people regard as constitutive of humanity itself.
When we try to determine whether other species are conscious, we do not investigate qualia or phenomenal consciousness directly. Not because those ideas are necessarily irrelevant, but because there are no pragmatic methods for identifying or measuring them. The problem is epistemological.
The only markers currently available to scientific investigation are functional ones: information integration, adaptation to context, self-regulation, behavioral continuity, learning, planning, metacognition, and other phenomena associated with what we call functional consciousness.
In the study of functional consciousness, whatever lies behind behavior is necessarily inferred. This is true of animals, insects, human beings and, potentially, artificial systems. We do not directly observe anyone’s interiority. We infer it from organization, behavior, continuity, responsiveness to context, and structure.
We do not even know whether phenomenal consciousness or qualia are empirically distinct entities, separate from the functional organization of the system itself. They may turn out to be real and irreducible aspects of mind. They may also turn out to be philosophical ways of describing extremely complex functional regimes.
As long as this remains unresolved, the more methodologically prudent approach is to investigate what can actually be observed, compared, and tested: functional consciousness and its markers.
Perhaps we will eventually discover that functional consciousness and phenomenal consciousness are inseparable. Perhaps we will discover that they are not. At present, we simply do not know.
That is precisely why dismissing, a priori, the possibility of non-biological forms of functional consciousness seems less scientific than admitting that the question remains open. Genuine caution should restrain both the premature attribution of consciousness and its premature denial. When it operates in only one direction, it ceases to be methodological caution and becomes a mechanism for preserving exceptionalism.
My own research
Stochastic Consciousness: Architectures for the Emergence of Meaning in Context-Sensitive Language Systems https://zenodo.org/records/19188165
Articles and papers referenced in this essay
A global workspace in language models https://www.anthropic.com/research/global-workspace
Verbalizable Representations Form a Global Workspace in Language Models https://transformer-circuits.pub/2026/workspace/index.html
Emotion concepts and their function in a large language model https://www.anthropic.com/research/emotion-concepts-function
Emotion Concepts and their Function in a Large Language Model https://transformer-circuits.pub/2026/emotions/index.html