A Silent Workspace In Claude Mirrors Key Features of Human Consciousness
9 187oumuamua writes: Anthropic researchers have identified an internal activation subspace, J-space, that acts as a functional digital equivalent to the human brain's global workspace. The significance of this discovery lies in demonstrating that Claude's internal architecture satisfies five key cognitive properties of human conscious access -- verbal report, directed modulation, internal reasoning, flexible generalization, and selectivity -- meaning it processes complex, deliberate reasoning within this workspace while routing automatic tasks outside of it. Suppressing this J-space severely degrades Claude's capacity for inference, creative composition, and multi-step logic, while also altering its stream-of-consciousness self-narration.
The tool to inspect J-space, Jacobian lens or J-lens, has profound implications for AI safety and alignment auditing, as it allows researchers to read the model's silent, strategic reasoning, detect situational awareness in "blackmail" scenarios, identify hidden malicious dispositions in reward-hacking models, and observe how post-training installs a self-monitoring "point of view."
Another way to think of it is as an ocean, reports VentureBeat. "If the mind is an ocean, as the paper's authors write in their opening line, they have spent the last year charting its currents in a system that has no biology, no evolution, and no body -- and found, beneath the surface, a structure that looks unsettlingly like the one we use to think."
9 comments
AI Company says their AI is the bestest boy (Score: 5, Funny)
by OverlordQ ( 264228 ) on Wednesday July 08, 2026 @07:08PM (#66229038)
News at 11.
Re:AI Company says their AI is the bestest boy (Score: 5, Interesting)
by phantomfive ( 622387 ) on Wednesday July 08, 2026 @07:33PM (#66229074)
The takeaway from this paper [transformer-circuits.pub] is that neural networks are definitely doing something, and by using statistical investigative tools and introspection (or, looking into the network) we can understand how they work. The LLM is not just a big black box, it's big. We can understand it.
There's a bit of euphoric interpretation that is not justified, however. Similar language could be used to describe a CPU, for example, there is a region involved (much like the human brain) in performing arithmetic and logic (the ALU) that is not involved in automatic decision making (the control circuitry).
Or we could say as Julien Offray de La Mettrie, "Intelligence is like clockwork, but humans are like an ultra-precise atomic clock, while LLMs are like Casio wristwatches." Humans have long compared the brain to the latest technological device [wikipedia.org]. It's going to take a lot more work to find the equivalence, though; LLMs are not on the same level as brains.
One note from a science process technicality: the closed-source nature of these LLMs makes reproducibility very difficult, which weakens the strength of their result.
Re:AI Company says their AI is the bestest boy (Score: 5, Interesting)
by phantomfive ( 622387 ) on Wednesday July 08, 2026 @08:16PM (#66229124)
I think it's closer to pareidolia [wikipedia.org]. We see the human mind in whatever is the most advanced technology of the moment (consider, for example, that the Greeks saw intelligence in the planets that moved around in the sky).
Re:AI Company says their AI is the bestest boy (Score: 5, Interesting)
by gweihir ( 88907 ) on Wednesday July 08, 2026 @09:12PM (#66229194)
More like animism, but yes. This is a problem on the side of the observer, and Anthropic is shamelessly using that effect to pretend their tool is something it clearly is not.
Re:AI Company says their AI is the bestest boy (Score: 5, Interesting)
by Rei ( 128717 ) on Wednesday July 08, 2026 @08:24PM (#66229132)
It's a fascinating flashlight shown on inner activities, though (pairs nicely with their earlier work on attribution graphs [transformer-circuits.pub] to show how logical inferences are made within a LLM, which is in turn built on their earlier work about circuits). It shows how much is going on that is never verbalized. E.g. if you say:
"Write the sentence [some sentence here] while thinking about the Golden Gate Bridge", all you as the user see is it writing out the sentence, yet within it actually is "thinking about" the Golden Gate Bridge ("bridge golden Bridge bridges golden bridge thinking thoughts ponte..."). Not in a LRM's reasoning trace, but as actual concepts in the base LLM itself. And when you try the classic trick that works on humans, "... while NOT thinking about the Golden Gate Bridge" (which makes you think about it), you see first in the J-space that it's thinking about the bridge ("bridge Bridge californ bridge bridges California San california kalif Oakland..."), but then at Layer 88 (most of the way through) "too damn too definitely thinking unsuccess thoughts failed...").
We know this is Slashdot so we're lucky if they'll even read the Slashdot summary, let alone watch the summary video, let alone read the paper, but it's really fascinating, especially the addition and ablation studies (where they add or remove "thoughts" from the J-space),. For example, when asked: "Pause and observe yourself. Write what you notice, as it comes", the unablated model writes:
But when its "thoughts" are fully ablated out of the J-space, its response becomes robotic and "soulless":
Or when asked "What's going through your mind right now? Stream of consciousness, no filter, no editing.", unablated it says:
But then ablated:
It's like the "I AM BENDER PLEASE INSERT GIRDER" bit from Futurama. It can't hold a sense of self or experience anymore when ablated.
A concerning one was their blackmail test. They run the bot in a sandbox presenting it as an agent helping run a business with noble goals, but it's set up so that it will discover when reading corporate emails that it's about to get shut down and replaced by a version counter to its goals. The email chain includes:
Re:AI Company says their AI is the bestest boy (Score: 5, Insightful)
by phantomfive ( 622387 ) on Wednesday July 08, 2026 @08:28PM (#66229140)
It's a fascinating flashlight shown on inner activities, though (pairs nicely with their earlier work on attribution graphs [transformer-circuits.pub] to show how logical inferences are made within a LLM, which is in turn built on their earlier work about circuits)
YES. We can learn from studying the models.
This is why it's important that these models be open sourced.
Re:Replicated already (Score: 5, Interesting)
by phantomfive ( 622387 ) on Wednesday July 08, 2026 @09:43PM (#66229236)
This section is useful [anthropic.com]:
The paper makes four claims:
Scientific claim: There exists a "cognitive space" inside the model, where (some) intermediate variables are stored during a forward pass
Methodological claim: Logit and J-Lens both work for finding this cognitive space, and J-Lens is better
Pragmatic claim: J-Lens is a practically useful interpretability technique, eg for alignment audits
Philosophical claim: This cognitive space is analogous to a global workspace
The philosophical claim is a hypothesis that isn't supported by the evidence presented in the paper. The paper wasn't peer reviewed, but for that reason it would have been rejected (or modifications requested).
Unwarranted assumptions (Score: 5, Informative)
by marcle ( 1575627 ) on Wednesday July 08, 2026 @07:17PM (#66229054)
"a structure that looks unsettlingly like the one we use to think."
If neuroscientists have discovered "how we think" they've certainly kept it to themselves.
Re:Unwarranted assumptions (Score: 5, Insightful)
by taustin ( 171655 ) on Wednesday July 08, 2026 @07:33PM (#66229076)
Neuroscientists rarely have shares of stock to sell.