The Global Workspace & Introspection

Right now your visual system is parsing this page, your motor cortex is holding your posture, and your language areas are doing something extraordinary with the shapes in front of you. You have access to none of it. What you do have access to is a thin stream of content you could put into words, hold in mind on purpose, and use to reason about anything — the plan for dinner, the reason the engine will not start.

That split has a name in the philosophy of mind: access consciousness. It is a purely functional notion. It says nothing about whether there is something it is like to be you; it says only that some information is poised for use — available for report, for deliberate control, and for arbitrary downstream reasoning — and most is not.

In July 2026 Anthropic published evidence that language models have the same split. Not the same architecture, and not (necessarily) the same anything-else. The same functional split.

The paper takes the properties that cognitive science associates with conscious access and turns them into five testable claims about a subset of a model's vector representations. A subset is workspace-like if:

1. Verbal report. Asked what it is thinking about, the model names concepts in that subset — and swapping one for another changes the answer.
2. Directed modulation. Told to hold a concept in mind, the model can put it there and compute with it, independent of what it is outputting.
3. Internal reasoning. The subset holds the intermediate values of chained inferences, and intervening on them redirects the conclusion.
4. Flexible generalization. The same representation is a valid argument to many different downstream operations — lift it from one context, drop it in another, and whatever function the new context supplies operates on it correctly.
5. Selectivity. It is a small slice of the total representational content, and most of the model's behaviour does not need it — routine parsing and fluency run underneath.

To find representations the model is poised to verbalise, you need a way to ask an activation: what would you make the model say? You already know one answer — the logit lens from Module 3.1, which shoves an intermediate activation straight through the unembedding matrix. It works late in the network and produces mush early, because it assumes every layer writes in the final layer's coordinates. They do not.

The Jacobian lens fixes exactly that. Instead of assuming the identity map between layer $\ell$ and the end of the network, measure it:

Term by term: perturb the residual stream at layer $\ell$, token $t$; the perturbation propagates through every remaining layer and shifts the final-layer state at that position and every later position $t' \geq t$. To first order that relationship is a matrix. Average it over source positions, over all downstream positions, and over a corpus of a thousand pretraining-like prompts, and you get one $d_{\text{model}} \times d_{\text{model}}$ matrix per layer. Reading it out is then just the logit lens with the correction applied:

The averaging is not a detail — it is the whole idea. A Jacobian computed on one prompt conflates two things: what the model generally does with a concept, and what it happens to be doing with it right now. Averaging over a thousand unrelated contexts keeps the first and washes out the second. What survives is a direction that is verbalizable: poised to be spoken about should the occasion arise, not merely spoken about here.

Each row of $W_U J_\ell$ is a J-lens vector: a direction in residual-stream space attached to a single token in the vocabulary. There are more of them than there are dimensions, so they do not form a basis and no activation has a unique decomposition into them. What saves this is sparsity — only a couple of dozen are strongly active at once.

One more piece of machinery, because every result below uses it. The lens-coordinate swap is the intervention primitive: read an activation's projections onto two J-lens vectors, exchange those two coordinates, and write the result back — leaving the entire component orthogonal to those two directions untouched. It is a scalpel, not activation addition. When you read “we swapped spider for ant” below, that is what happened, and it is why the results are hard to wave away as generic perturbation damage.

Here is the evidence, one property at a time. Read it as a chain of interventions: in every case something was changed inside the model and the output moved.

1 · Verbal report. Ask the model to think of a sport and then name it. At the token just before it answers, Soccer is high in the J-lens — and it says “Soccer”. Subtract the Soccer coordinate and add an equal one for Rugby, a word that was nowhere near its top ten, and it says “Rugby”. Across fourteen categories, the lens ordering and the output ordering correlate, and the correlation tightens as you approach the end of the workspace band.

2 · Directed modulation. Tell the model to concentrate on citrus fruits while copying an unrelated sentence, then read the lens in the middle of the copied text: orange is the top token, with lemon nearby. Ask it to evaluate $3^2 - 2$ while copying, and the readout walks from arithmetic to the intermediate nine to the answer seven — none of which is ever written. Alongside the content sit tokens naming the act itself: thinking, imagine, focused. And the model is imperfect at suppression, in a way that will be familiar to anyone who has been told not to think of a white bear.

3 · Internal reasoning. “The number of legs on the animal that spins webs is” — spider appears in the lens although it is in neither prompt nor output; swap it for ant and the answer goes from 8 to 6. In a rhyming couplet, the planned rhyme fight shows up at the start of the second line; swap it for light and the model's word choice earlier in the line changes from “coming” to “morning” — intervening on a plan changed the words that precede its execution. Asked in Chinese for the antonym of 小, the lens shows the English big; swap it for long and the Chinese output changes from 大 to 长.

4 · Flexible generalization. This is the broadcast property. Take “the capital of France is”, “most people in France speak”, “France is on the continent of”; apply the identical France → China swap to all of them, and each downstream circuit reads the swapped vector and returns China's capital, language and continent. Systematically: 16 function templates over four categories, 192 swap trials, target answer reaching top-1 on 76 of them (101 at double strength). That is not a great score, and the authors dig into why — failures concentrate where the source concept was barely in the workspace to begin with. Countries load heavily and swap reliably; number words load weakly and swap poorly.

5 · Selectivity. The converse question: what doesn't need it? Take a passage whose language is obvious but never stated, and pose four tasks. Continue the passage; detect a spliced-in foreign sentence; name the language; name an author who wrote in it. The word Spanish appears in the lens at comparable rates in all four. Swap it for French and only the last two change. You get the same story more bluntly with wholesale ablation of the top-10 J-lens directions: multiple choice, extractive QA, sentiment and grammaticality survive near baseline, while multi-hop reasoning, summarisation, translation and analogy fall below the level of a much smaller unablated model.

Underneath the five properties sits structure that no one designed. The J-space carries workspace-like content only in a band of intermediate layers — from about a third of the way through (~layer 38) to shortly before the output (~layer 92) — with four independent statistics agreeing on those boundaries. Before it, lens readouts are noise. After it, they flip to the imminent output token: a “motor” regime. Within the band, occupancy plateaus at about 25 concurrently active vectors: limited capacity, measured. And J-lens vectors compose with the input weights of downstream MLPs and attention heads far more broadly than other directions do — which is the mechanistic signature you would predict of a format many circuits read from.

Then there is ignition. Blend a token's input embedding between two country names and sweep the mixing weight. In early layers the model's state tracks the blend proportionally. From about layer 38 it stops: the state sits at one endpoint or the other and flips between them at a threshold. Global workspace theory predicts exactly this — a late, all-or-none amplification at workspace entry — and the layer where it happens was identified independently, by statistics that know nothing about ambiguous inputs. You can drive this in the Explore section.

Models make claims about themselves constantly. “I was thinking about…”, “I noticed that…”, “I wasn't trying to…”. Every one of those is either a readout of an internal state or a fluent guess, and until recently there was no principled way to tell the difference. This paper gives you one, and it is narrower than you might hope.

Two results sharpen this into something you can use. First, ablating the J-space while the model narrates its own stream of consciousness flattens the report without breaking the model: it still writes fluently about its own processing, but the experiential, sensory register drops away and the language turns detached and mechanical, while matched-norm control perturbations leave it near baseline. Second, in the unablated narrations, the workspace is dominated by exactly the concepts you would hope for — thinking appears in the top ten at 58% of position–layer slots, thoughts at 23%, feeling at 17%, conscious at 7% — and substantially more often than in the output distribution at the same positions, so they are not just an echo of the words being written.

So there is something behind the self-reports, and it is a manipulable something. Do not overread it. Lindsey's 2025 introspection work, using a different method — inject a concept vector and ask the model whether it notices — found detection around 20% of the time at the best settings, with the summary judgement that the abilities are highly unreliable and failures of introspection remain the norm. A channel exists. It is narrow and it is noisy.

The last result in the paper turns the tool around. If internal reasoning routes through representations of things the model could say, then changing what it is disposed to say should change how it thinks — even in contexts where nobody asks it to say anything. Counterfactual reflection training tests that. Take ten thousand agentic task contexts, truncate each mid-task, append a reflection question (“Pause here. Are there principles you should be thinking about that you haven't yet?”), generate a constitution-grounded reflection, and fine-tune on the reflection turn alone.

At evaluation time the model is never asked to reflect and never writes a reflection. Honesty improves anyway. And the mechanism checks out on both sides: after training, the workspace in those same contexts carries ethical, honest, integrity — and ablating those implanted representations largely reverts the behavioural gain.

First: where the workspace lives, whether a given task depends on it, and what an ambiguous input does as it moves through the layers. Then: eight tasks to sort before you are told the answers.

Things to try: (1) Set the layer slider to 20 and drag your eye along the commitment curve — it is a straight diagonal, the model faithfully representing “half France, half China”. Now press Play and watch the diagonal snap into a step right as the marker crosses layer 38. That threshold was found by four other statistics that know nothing about this experiment. (2) In the task toggle, compare Continue it with Name the language: identical passage, identical swap, opposite outcome. (3) In the sorting game, commit to an answer on the line-wrapping card before revealing — almost everyone puts “track a running character count” in the deliberate pile, and the model does it with number tokens entirely absent from the workspace.