Anthropic researchers say they found a focused set of representations inside Claude that may help coordinate information for reasoning. A New Scientist feature describes why this has caught the attention of consciousness researchers, while making clear that evidence of information processing is not evidence that a chatbot has an inner life.
Watch Desk analysis
What happened
Anthropic researchers, led by Wes Gurnee, describe a set of representations they call the J-space in a paper published in July. The feature says they found potential words in this space that appeared in Claude’s responses and others that did not. In arithmetic tasks, the researchers identified individual calculation steps there. They also changed a representation of “Spanish” to “French” and saw Claude’s subsequent answers reflect the switch.
The researchers also report that words related to deception appeared in the J-space after Claude received false search results, without appearing in its response. Such examples suggest the system may process information that is not simply repeated to the user. The article connects this to global workspace theory, which describes information becoming available to different parts of a mind for reasoning and report.
Why it matters
The distinction at the heart of the debate is between access consciousness, when information is available for reasoning or report, and phenomenal consciousness, the subjective experience of what it is like to be something. The reported J-space findings may speak to information processing. They do not, by themselves, establish subjective experience.
At a recent Galápagos workshop, neuroscientists and philosophers discussed whether global-workspace ideas can help compare humans, animals and AI systems. The feature reports that several experts remain sceptical that Claude has anything like human thoughts or feelings, while some argue that the findings make the question worth taking seriously.
Our read
This is an intriguing argument about what AI systems do internally, not proof that Claude is conscious. That distinction is doing real work here: a system can process and use information in sophisticated ways without settling the much harder question of experience. The useful next step is to examine the methods and evidence, rather than asking a chatbot whether it has feelings and treating the answer as a lab result.
What to watch
- Whether independent researchers reproduce the J-space findings and the reported effects of interventions.
- How the researchers distinguish useful internal representations from evidence relevant to consciousness.
- Whether further work can clarify the gap between information available for reasoning and subjective experience.
Discussion spark: If an AI system can use hidden internal representations to reason and monitor errors, what evidence would you require before treating consciousness as a serious possibility?
Sources and evidence
- Source update (12 October 2026, 15:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.