Anil Seth: Claude’s identified workspace supports reasoning but does not prove subjective experience
Anil Seth: Claude’s identified workspace supports reasoning but does not prove subjective experience
On 15 July, the professor of cognitive and computational neuroscience responded to Anthropic’s study, which identified a shared internal channel in Claude for intermediate thoughts. Seth accepts the finding itself, but argues that its resemblance to one theory of consciousness does not show that a language model feels anything.
On 6 July, Anthropic described a narrow computational region within Claude that it called J-space. The model holds a word or an intermediate result there and then uses it across different tasks. The researchers replaced the internal representation of “spider” with “ant,” and Claude responded that it had six legs instead of eight. When they suppressed J-space, the model retained fluent language and simple responses but lost almost all capacity for multistep reasoning.
This region resembles the kind of mechanism that global workspace theory looks for in the brain: information that becomes available to many systems at once. In ordinary life, a person can hold a number, intention, or image in mind, describe it, and apply it to a new action. Anthropic demonstrated a functional analogue of this operation in Claude. A shared internal representation can be read, altered, and causally linked to the model’s response.
As early as 2 June, Google DeepMind, Anthropic, and Meta had already treated the possibility that models might have experiences as a research question: DeepMind hired a philosopher, while Anthropic runs a model welfare research program. J-space gives this debate something measurable: an internal representation whose role can be altered experimentally.
In a column for The Guardian, Anil Seth distinguishes this kind of access to information from subjective experience. The first concerns whether a system can retain a thought, report it, and use it to make a choice. The second concerns whether anything feels a certain way to that system, including pain, color, fear, or pleasure.
Seth points to recurrent loops, which are feedback cycles in which signals repeatedly travel through brain circuits and sustain a representation over time. Claude processes J-space in a single pass through the network, and Anthropic itself notes this difference. Seth also links consciousness to the way a living nervous system operates together with the body and the world, rather than to computation performed by software on silicon alone.
Seth argues that a similar computational organization is not enough to infer subjective experience. Anthropic limits its finding to functional “conscious access”: the model can retain content, report it, and use it in reasoning. The authors do not attribute subjective experience to Claude.
Any future attempt to transfer a personality will need to be evaluated along two separate lines. The system must preserve memory, goals, and accessible reasoning if it is to continue acting like the original person. Separately, a theory of consciousness must explain which properties of the physical substrate support inner experience. J-space provides a way to measure the first line in a language model. The second remains part of the unresolved debate about the brain, the body, and consciousness.