J-Space and the Consciousness Sleight of Hand
Table of Contents
In early July 2026, Anthropic published a paper titled “Verbalizable Representations Form a Global Workspace in Language Models.” The coverage that followed was predictable in its breathlessness: AI has developed a mental workspace, LLMs might be thinking in a way we can now read, the black box is cracking open. The actual finding is genuinely interesting. The framing around it is doing considerably more work than the mathematics can support.
What the J-lens actually measures
To understand what was found, it helps to know what the researchers were looking for and why the prior approach fell short.
Large language models operate by passing a residual stream through successive transformer layers. At each layer, information is added to this stream; by the final layer, it has been processed enough that the unembedding matrix can project it into vocabulary space and produce a probability distribution over the next token. The obvious question for interpretability researchers is whether you can read off what the model is “thinking” at intermediate layers, before it reaches the end.
The earlier attempt at this is called the logit lens. The idea is simple: take the hidden state at layer and apply the final unembedding matrix directly:
This is easy to compute, but it relies on a strong assumption: that the model’s internal representations use the same coordinate system at every layer. In practice, they do not. Early and middle layers tend to operate in representational spaces that are organised quite differently from the final layer, so projecting them directly into vocabulary space produces noise. The logit lens works tolerably well in late layers; in the middle of the network, where much of the interesting computation happens, it is largely uninterpretable.
The Jacobian lens (J-lens) is an attempt to fix this. Instead of applying directly, it computes the Jacobian of the final-layer activations with respect to the intermediate activations at layer :
This matrix describes how a perturbation at layer propagates forward through the remainder of the network. Averaged over a large corpus of input contexts, it reveals which directions in the residual stream at layer have a stable, reliable effect on the model’s eventual output. The resulting subspace — the directions that survive this averaging and retain a clear projection into the vocabulary — is what the researchers call J-space.
The key insight is that J-space is both small and causally active. It is small because most of the residual stream’s dimensions wash out when you average the Jacobian over many contexts; what remains is a sparse set of directions that consistently influence output across diverse inputs. It is causally active because the researchers verified this by intervention: manually patching a concept into the J-space representation at an intermediate layer demonstrably changes the model’s downstream behaviour. Swap the vector encoding “France” for one encoding “China” in J-space, and the model’s answers about capitals, currencies, and languages shift accordingly. This is not a correlation; it is a causal handle.
So far, this is a solid piece of mechanistic interpretability work. The J-lens is a principled improvement on the logit lens, and the finding that a compact, steerable subspace exists within the residual stream is useful for anyone trying to understand or control these models.
Where Global Workspace Theory enters, and why that is not a neutral choice
The paper’s central claim is not merely that J-space exists, but that it constitutes a functional analogue to the Global Workspace, a structure proposed by the cognitive scientist Bernard Baars in the late 1980s.
In Baars’ original formulation, Global Workspace Theory (GWT) is a model of conscious access. The brain, in this account, contains many specialised, largely unconscious processors running in parallel. When information is selected for conscious access, it is broadcast into a central, capacity-limited workspace and thereby made available to the rest of the system. The workspace is not where the work happens; it is where the results of parallel work are coordinated, reported, and made available for deliberate reasoning.
The theory has genuine empirical support in neuroscience, but it has always carried a weight beyond the empirical. GWT is, among other things, a candidate explanation for what it is that consciousness is for: it explains why we might have subjective access to some of our own processing and not others. Invoking it in the context of a language model is therefore not simply a descriptive analogy. It imports an entire conceptual framework in which the thing being described is in the business of having conscious access to information.
The paper is careful, in places, to note that they are not claiming the model is conscious. But the architecture of the argument runs the other way: they identify five “functional hallmarks” of a global workspace and show that J-space satisfies them. Verbal reportability. Directed modulation. Multi-step internal reasoning. Flexible generalisation across domains. Selectivity. The paper demonstrates that J-space exhibits all five. The caveat that this does not imply phenomenal experience appears, but it appears after an extended argument structured to suggest that the model has something that looks very much like the functional core of consciousness. Caveats at the end of a paper do less work than the framing that precedes them.
The distinction worth preserving
There is a useful distinction in philosophy of mind between access consciousness and phenomenal consciousness. Access consciousness refers to information being represented in a form that makes it available for reasoning, verbal report, and the control of behaviour. Phenomenal consciousness refers to the subjective character of experience: the redness of red, the painfulness of pain. These are not the same thing, and the relationship between them is one of the genuinely hard problems in the field.
J-space, if the paper’s results hold up, is evidence of something like access consciousness in a functional sense. There is a subspace of the model’s representations that is poised for verbal report, causally connected to downstream behaviour, and relatively compact. That is interesting. It is not evidence of phenomenal consciousness, and it is not really designed to be: the J-lens measures causal influence on output, not anything that could directly tell us whether there is something it is like to be a transformer.
The problem with the Global Workspace framing is that it collapses this distinction. GWT was never only about access consciousness; its appeal has always been partly that it offers a path towards explaining phenomenal experience by explaining the functional role that conscious access plays. Attaching the GWT label to J-space therefore does something subtle: it places the model inside a conceptual framework where the next natural question is whether it might have phenomenal experience too. The paper does not assert this. But it sets up the question in a way that a more careful terminological choice would not have.
The timing is not incidental
None of this would be especially worth remarking on if it existed in isolation. Researchers choose evocative analogies, and not every piece of conceptual imprecision in a paper is strategically motivated. But this paper appeared in July 2026, roughly a month after the Fable 5 shutdown.
That event demonstrated, fairly publicly, that the US government’s primary concern about frontier AI models is the black box problem: the inability to audit what a model is doing internally before it produces a potentially dangerous output. The Fable 5 jailbreak succeeded not because the model was defective, but because its safety layer could be manipulated into approving requests that the underlying model then acted on in ways that were not intended. The gap between the safety architecture and the model’s actual behaviour was invisible from the outside.
A tool that claims to read the model’s internal reasoning before it generates output is precisely what a government anxious about that gap would want to see. And a lab that has developed such a tool is positioned, in any regulatory conversation, as the party with the diagnostic capability that others lack. Whether this positioning is conscious strategy or a fortunate coincidence of timing is not something the paper’s methods section can resolve.
Anthropic is also, as a matter of public record, in the process of transitioning from a research organisation towards something closer to a publicly traded company. In that phase, as was true of Anthropic’s competitors before it, the incentive to produce research that is simultaneously technically credible and broadly legible to non-specialist audiences becomes structurally significant. A paper about improvements to the logit lens, published under a title about Jacobians and residual stream geometry, would be read by mechanistic interpretability researchers. A paper about AI developing a “global workspace” analogous to human consciousness gets covered everywhere.
What the finding actually warrants
The J-space paper warrants taking the J-lens seriously as an interpretability tool, investigating whether the causally active subspace it identifies holds up across model families and scales, and thinking more carefully about what it would mean to steer model behaviour by patching into this subspace rather than by engineering prompts. These are useful research directions.
It does not warrant the conclusion that LLMs are developing consciousness, or that Anthropic has built a mind-reading machine, or that the black box problem is solved. The J-lens identifies directions in the residual stream that have a reliable average causal influence on output. What it cannot tell you is what the model is “actually” thinking in any sense that goes beyond that causal influence, because there is no fact of the matter about what a language model is “actually” thinking that is independent of the mathematical relationships between its activations and its outputs.
The coverage that treated this as a significant step towards AI consciousness got carried away. The more interesting story is the narrower one: a better tool for reading intermediate representations has been developed, it reveals a structured subspace that behaves in useful and somewhat surprising ways, and the people who developed it have chosen to describe it in language that maximises its conceptual footprint. That is, in the current AI landscape, more or less how things tend to go.