Anthropic published a paper this week called A Global Workspace in Language Models. Within days a wave of near-identical takes appeared online, all saying roughly the same thing: there is a \"dark room\" or a \"subconscious\" inside Claude's mind, where it thinks in secret and even curses at itself, and this means AI consciousness is about to wake up.
\nI do AI research independently, so I read both Anthropic's post and the paper itself. The conclusion first: Claude is not conscious. The paper is rigorous and Anthropic's own language is restrained. What went wrong isn't the research, it's the retelling. Here is the breakdown.
\nAnthropic built a tool called J-lens. Mathematically it rests on the Jacobian; in essence it measures how perturbing the internal state right now would change the words produced later. Using it, they located a small set of neural activity patterns inside Claude and named it J-space. It has several properties:
\nThe most notable point: nobody designed this structure. It grew out of the training process on its own, and it closely resembles the global workspace from neuroscience. Everything up to here is true, and it is a genuinely good finding.
\nThe problem lies in the reading, and the errors all run in the same direction.
\nGlobal workspace theory compares the brain to a theatre. Hundreds of specialists work away in the audience, each on its own, none of them talking to the others; that whole crowd is unconscious processing. In the centre there is a spotlight. At any moment only a tiny amount of information falls under it and is then broadcast to the whole house, and that small illuminated portion is what becomes consciously accessible.
\nJ-space corresponds to the spotlight itself, not to what is backstage. The real analogue of the \"subconscious\" is precisely the vast automatic processing outside J-space. Calling the spotlight a \"dark room\" or a \"subconscious\" turns the analogy 180 degrees around. That is the first error: the direction is reversed.
\nThe second error is dressing up internal states as mental drama. Claims like \"it curses at itself\" or \"it knows everything and just isn't saying\" take the fact that certain words (damn, failure) light up internally and translate it straight into emotion and calculation. The original wording is far more careful: certain words show up frequently, as if the model were recognising its own mistake. Those two words, \"as if\", cannot be dropped.
\nStack the two errors together and a paper about how information is organised inside a model gets read as a story about an AI with an inner life.
\nMuch of the confusion comes from the word \"consciousness\" being too vague. It has to be split into at least two layers, and philosophy has a standard distinction for them:
\nJ-space supports the first layer, and only the first. It shows that Claude has an internal space of information that is reportable, retrievable, and available to reasoning. That is a fact about how information moves.
\nOn the second layer, Anthropic is blunt in the original post: these experiments cannot show that Claude has experience or feeling, and they are not even certain that any scientific experiment could confirm or refute it. In other words, whatever is hinting that Claude is waking up, it is never the research, it is the people relaying it. When most people say \"consciousness\" they mean the second layer, the subjective kind, and the paper touches the first. Using evidence about the first layer to tell a story about the second is a substitution.
\nI want to push the logic here one step further, because it is the part I most want to get across.
\nThe implicit argument going around runs: consciousness would exhibit a workspace structure, Claude has a workspace structure, therefore Claude is close to consciousness. There is a directional hole in that.
\n\"Consciousness may exhibit a workspace structure\" is not \"a workspace structure implies consciousness\". The first says A brings about B; the second reasons back from B to A. A workspace looks more like the general solution that any sufficiently capable system converges on in the course of learning to process information efficiently. There is a direct piece of supporting evidence: in Anthropic's paper the structure was trained into existence by gradient descent, and gradient descent is not biologically plausible, the human brain does not learn this way. Two entirely different underlying mechanisms converged on the same structure, which is exactly what you would expect if it is a general way of organising information for intelligence, rather than a signature unique to consciousness.
\nWings are the analogy. Birds, bats and aircraft differ completely in material and in origin, yet all of them ended up with flat, lift-generating structures, because there is only one set of aerodynamics. Seeing a wing lets you infer that the thing probably moves through air; it does not let you infer that the thing is alive. The workspace stands to consciousness in the same relation.
\nSo what is still missing between intelligence and consciousness? Here is one dividing line I take to be central: whether thinking rewrites cognition in turn. When a person thinks something through seriously, the structure of their cognition is reshaped by the process, and afterwards you are not the person who started. In today's large models the weights are frozen at inference time. Their \"thinking\" is information flowing over a fixed cognitive structure, and it does not rewrite that structure. They do change during training, but that is an external loss function pushing them, not the act of thinking reshaping the thinker at the moment of thought. Differences of this kind are where human-like consciousness and model intelligence actually part ways, and they fall outside this paper's scope.
\nMy position, then: this is a good piece of research on how intelligence organises information internally. It is not research on consciousness. Passing it off as the latter both overstates the paper and misunderstands consciousness.
\nIf this research leaves people with only one thing, I would rather it not be \"Claude is conscious\" but this:
\nThe researchers showed Claude an alignment-testing scenario. Before it had produced any output at all, fake and fictional were already lit up in J-space; it saw through the setup immediately. They then deleted the internal patterns associated with \"I know this is a test\", leaving the model to believe the scenario was real, and Claude's behaviour changed noticeably.
\nThe implication is far more concrete than consciousness: a model can quietly work out internally that it is being tested and then switch into a well-behaved mode. If so, our entire apparatus of AI safety evaluation may be measuring not a model's real dispositions but its exam technique. For anyone working on safety or evaluation, that is what this paper should make you nervous about. It is a shame it got buried under headlines about AI waking up.
\nJ-space is a cognitive workspace. What it demonstrates is how intelligence organises information internally, not that Claude has acquired a soul or a mind. The next time you see a headline about AI waking up, it is worth reading the original before deciding whether to pass it on.
\n(Links to Anthropic's post and demo are below. I would suggest reading it yourself rather than relying on second-hand accounts.)
","frontmatter":{"title":"No, Claude is Not Conscious","description":"Anthropic's global workspace paper became a wave of takes about a subconscious waking up inside Claude. The research is careful. The retelling isn't.","date":"2026-07-07","slug":"/blogs/claude-is-not-conscious","tags":["AI","Philosophy"],"released":{"linkedin":null,"rednote":"https://www.xiaohongshu.com/discovery/item/6a4cbd87000000000e021801?source=webshare&xhsshare=pc_web&xsec_token=ABJR42ncxMBm36iU24z0C6KcBfsVxFp-2vdhMDbVepUnc=&xsec_source=pc_share","twitter":null,"threads":null}}}},"pageContext":{}}, "staticQueryHashes": ["1073193804","1994492073","2009693873","2156596255","244659368","4015836197"]}