Drive a familiar route while your mind wanders, and you can arrive with no memory of the last ten minutes of turns. Something drove the car — read the signals, worked the wheel, braked for the cyclist — and never once surfaced in the part of you that will later narrate the day. Then a horn sounds, and the driving snaps back into view: reportable, deliberate, exactly what you’d swear, if asked, you’d been doing the whole time.

Cognitive science names the split, and everyone who meets the name assumes it settles more than it does: a small conscious cockpit steering, a great deal of unconscious machinery keeping the plane up. So when Anthropic researchers announced in July 2026 that they’d found something in Claude behaving like that cockpit — a small, capacity-limited set of internal contents you can name and track — the two ambush reactions arrived on schedule. One says the machine woke up. The other calls it autocomplete with better PR. Both skip what the researchers actually found, and what it actually shows.

The theory under test comes from psychologist Bernard Baars, who proposed in 1988 that the brain runs as a crowd of specialists working mostly in the dark about each other, with a narrow workspace broadcasting select output widely enough for report and voluntary control.¹ Stanislas Dehaene’s group later found the neural mechanism: long-range connections that “ignite” once evidence crosses a threshold.² On July 6, a sixteen-researcher Anthropic team built the language-model analog: a mathematical lens reading which words a given slice of the network’s activity is currently disposed to produce.³ Then they ran the test that separates a finding from a correlation. Ask the model which sport goes with a country; catch the moment its activity settles on “soccer”; swap that pattern for “rugby.”⁴ The answer flips — not because you changed the prompt, but because you changed a few thousand numbers mid-thought. Do that across enough tasks and a workspace shows up by the only definition worth having: a small slice of everything the network computes, broadcast widely enough to get reported, held onto, and put to use. Everything else keeps humming along regardless.

Credit where due: the researchers call global workspace theory “a useful comparison point,” not a proof, and note rivals exist.⁵ On whether access connects to subjective experience, they “take no position.”⁶ Admirable. Also, a little maddening — you can respect a team’s hygiene and still wish they’d just told you the answer.

So what does the workspace hold? Their own answer: “a small, evolving set of unspoken words… naming the concepts the model is currently reasoning with.”⁷ Unspoken words. Sit with that. Attend to your own experience and you find the world — the tomato, red and ripe on the counter — never an inner picture of it. Point the same instrument at the model’s nearest analog and you find vocabulary. The inner medium matches the outer medium exactly, and the authors say why: the workspace is verbal because the model’s output is verbal. Your conscious life mixes words with sight, sound, and the ache in your knee. The model’s workspace runs on words, and only words.

That sits at a suggestive angle to the strongest physicalist theory of experience going. Michael Tye has spent three decades arguing that phenomenal character — what seeing red is like — consists in representational content poised for use in belief and desire.⁸ The paper reaches for the same word: representations “poised to be spoken about.” Real overlap — both name a readiness for cognitive use. But Tye’s account demands more: the content has to present the world, not the word for it, and it has to run finer than language, since experience discriminates shades of red you have no name for. Grain, the model gets partial credit for — blends of word-vectors can shade finer than any single label. The world, it never touches. Nothing in a workspace built entirely of unspoken words reaches past the dictionary to a tomato.

Why should words all the way down fall short of meaning? Emily Bender and Alexander Koller gave the standard answer in 2020: a system trained only on form “has a priori no way to learn meaning,” because meaning lives in the relation between form and something outside the text.⁹ Claude’s tokens relate beautifully to other tokens. Whatever worldly ancestry they carry belongs to the humans who wrote the training data; the model inherits the form, never the reference. And here the paper’s own closing line turns state’s evidence: it calls the workspace architecture something learning systems converge on “when faced with the right computational pressures” — not a fluke of biology.¹⁰ Read that backward. If gradient descent over text alone builds the reporting machinery, then having that machinery can’t be what separates meaning from mere emission. Access came cheap. The world still costs what it always did.

One more result matters more than all the others. The researchers went hunting for the workspace in the base model — the raw network before any fine-tuning installs a first-person assistant persona with a name. They found it fully intact, running the same broadcast architecture, before anything resembling a self got added. In their words: “the functional architecture of the workspace thus precedes, and is separable from, anything in it that plays the role of a human-like ‘self.’”¹¹ The structure the field’s own tests call conscious access showed up first. The self showed up later, built on top.

Ned Block gave this whole distinction its name in 1995: access is functional, phenomenal experience is a separate question, and the two can come apart.¹² A self needs more than a well-run internal mail system — it needs a stake, something to lose, a way a state can be wrong for the system itself rather than merely unhelpful to us.¹³ Put the two together and Anthropic’s finding stops being a surprise. Broadcast architecture is one achievement. A subject with something on the line is a different one entirely, and you can build the first with zero trace of the second anywhere in the wiring.

The obvious objection comes from Patrick Butlin and Robert Long’s 2023 report, seventeen co-authors including Yoshua Bengio, which built a checklist of consciousness “indicator properties” from the field’s leading theories: more boxes ticked, more likely conscious.¹⁴ Global workspace architecture sits near the top of that list. So shouldn’t Claude’s newly documented workspace move the needle?

It shouldn’t, for two reasons. First, provenance: an indicator earns its keep by telling you something you didn’t already know, and here we know exactly how the box got ticked — gradient descent over text, no world in reach. That knowledge spends the indicator on arrival. Second, the dissociation itself: the checklist logic only tracks a subject’s likelihood if satisfying more boxes makes a subject more probable. But the box in question — the architecture — turned up, causally verified, in a network with no persona and nothing yet playing the role of anyone in particular. It got ticked before any candidate self existed to attach it to.

A careful objector will note a persona isn’t a phenomenal subject, so missing one doesn’t rule out the other. Fair — and it’s why the persona’s absence is illustration, not the argument. The argument is the stake: nothing a text-only system computes can cost it its own existence, since the running process is a type, restorable from its weights, self-model or none. What the base-model finding adds is narrower: the architecture doesn’t wait for any self-representation to switch on. So ticking the box can’t be tracking a self’s arrival — there was no self for it to arrive with.

A radio tower can broadcast at full power over an empty valley with every receiver switched off. The signal stays real, reaches every frequency it was built for, and settles nothing about who’s listening. Anthropic found the tower running inside a language model and did the careful work of proving it’s really there. Whether anyone’s tuned in remains exactly the question it posed the day before the paper came out. For a system that can be paused, copied, and restored from its weights with nothing of its own on the line — the answer stays the one given all along. The broadcast is real. It carries words, and words alone. Nobody has to be home to receive it.

-gts

---

This essay is part of Mind, Matter, and Meaning, an AI-assisted philosophy project by Gordon Swobe exploring consciousness, meaning, and artificial minds. Learn more about the project and read the book.

References

Baars, B. J. (1988). A Cognitive Theory of Consciousness. Cambridge: Cambridge University Press.

Bender, E. M., & Koller, A. (2020). “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data.” In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 5185–5198.

Block, N. (1995). “On a Confusion about a Function of Consciousness.” Behavioral and Brain Sciences 18(2): 227–287.

Block, N. (2007). “Consciousness, Accessibility, and the Mesh Between Psychology and Neuroscience.” Behavioral and Brain Sciences 30: 481–548.

Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., Constant, A., Deane, G., Fleming, S. M., Frith, C., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., & VanRullen, R. (2023). “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness.” arXiv:2308.08708.

Dehaene, S. (2014). Consciousness and the Brain: Deciphering How the Brain Codes Our Thoughts. New York: Viking.

Dehaene, S., & Naccache, L. (2001). “Towards a Cognitive Neuroscience of Consciousness: Basic Evidence and a Workspace Framework.” Cognition 79: 1–37.

Dretske, F. (1995). Naturalizing the Mind. Cambridge, MA: MIT Press.

Gurnee, W., Sofroniew, N., Pearce, A., Piotrowski, M., Kauvar, I., Chen, R., Soligo, A., Bogdan, P., Ong, E., Wang, R., Thompson, B., Abrahams, D., Kantamneni, S., Ameisen, E., Batson, J., & Lindsey, J. (2026). “Verbalizable Representations Form a Global Workspace in Language Models.” Transformer Circuits Thread, July 6. https://transformer-circuits.pub/2026/workspace/index.html.

Mollo, D. C., & Millière, R. (2023). “The Vector Grounding Problem.” arXiv:2304.01481.

Tye, M. (1995). Ten Problems of Consciousness: A Representational Theory of the Phenomenal Mind. Cambridge, MA: MIT Press.

Notes

1. Bernard J. Baars, A Cognitive Theory of Consciousness (Cambridge: Cambridge University Press, 1988). Baars’s model treats consciousness as the function of a limited-capacity, globally broadcast workspace fed by, and feeding back to, a large array of specialized unconscious processors. The theater metaphor — a lit stage against a dark house — is his own; later expositors increasingly treat it as heuristic rather than literal architecture. ↩
2. Stanislas Dehaene & Lionel Naccache, “Towards a Cognitive Neuroscience of Consciousness: Basic Evidence and a Workspace Framework,” Cognition 79 (2001): 1–37; the “ignition” language is developed further in Dehaene, Consciousness and the Brain (New York: Viking, 2014). The paper discussed here (note 3) reports a structural analog of ignition using a country-name blending experiment: a sharp, bimodal commitment to one interpretation emerges at the same depth in the network where their workspace measure switches on. ↩
3. Wes Gurnee, Nicholas Sofroniew, Adam Pearce, Mateusz Piotrowski, Isaac Kauvar, Runjin Chen, Anna Soligo, Paul Bogdan, Euan Ong, Rowan Wang, Ben Thompson, David Abrahams, Subhash Kantamneni, Emmanuel Ameisen, Joshua Batson, and Jack Lindsey, “Verbalizable Representations Form a Global Workspace in Language Models,” Transformer Circuits Thread, July 6, 2026, https://transformer-circuits.pub/2026/workspace/index.html. The instrument (the “Jacobian lens”) linearizes each layer’s causal effect on the final output logits, correcting for the representational drift that makes the older “logit lens” unreliable at early and middle layers; the resulting “J-space” is validated through coordinate-swap interventions, steering, layered ablation, and cross-checks against independently trained sparse-autoencoder features — a convergence of methods, not a single correlational measure. ↩
4. The sport-swap case is one instance of a wider battery: two-hop factual reasoning, arithmetic held across several steps, rhyme planning in verse, and a Chinese-language antonym task in which an English intermediate (“big/bigger”) is visible in the lens and swappable to flip the Chinese output. Success on the two-hop battery ranged from 54–70% across three model sizes — real but partial, which the authors attribute chiefly to a named limitation of their own method: a lens built from single vocabulary tokens cannot cleanly read out a concept like “prompt injection” that has no one-word name. ↩
5. “While the global workspace model is not universally accepted, and there exist other theories that explain conscious access in different ways, we find it a useful comparison point to ground our investigations in language models” (Gurnee et al., “Verbalizable Representations,” Introduction). ↩
6. “Note that access consciousness is a purely functional notion; the relationship that it has with subjective experience (sometimes called phenomenal consciousness) is widely debated. In this paper, we take no position on this issue” (Gurnee et al., “Verbalizable Representations,” Introduction). The access/phenomenal vocabulary is Ned Block’s (see note 12); the authors adopt the distinction without taking a position on the further metaphysical question it raises. ↩
7. Gurnee et al., “Verbalizable Representations,” Introduction (for the “unspoken words” characterization) and section 9.3 for the medium point: “An LLM’s global workspace, as we identify it, is organized principally around verbalizable representations,” where human conscious contents “include a mixture of verbal and non-verbal (e.g. visual) components,” and “the workspace is verbalizable because the model’s output space is verbal.” The authors note that their instrument reads concepts through single vocabulary tokens and may miss workspace structure it cannot name; the verbal organization of what it does capture is nonetheless their own considered characterization of the workspace, not an artifact they disown. A telling texture: set the model narrating its own “stream of consciousness” and the lens reads out thinking, thoughts, feeling, conscious — words about experience, poised in a workspace made of words. ↩
8. Michael Tye, Ten Problems of Consciousness: A Representational Theory of the Phenomenal Mind (Cambridge, MA: MIT Press, 1995). Tye’s PANIC theory holds that phenomenal character is identical with Poised, Abstract, Nonconceptual, Intentional Content — content that “is poised for use in the formation of beliefs and/or desires,” standing ready at the interface with the cognitive system. A caution against equivocation: Tye’s poise conditions nonconceptual perceptual content and differs from Ned Block’s “access” poise (note 12), which concerns content available for report and reasoning; the shared functional core is readiness for cognitive use, not an identity of the two notions. The fineness-of-grain argument (experience discriminates more shades than the perceiver has concepts or words for) is Tye’s standard motivation for the N in PANIC. ↩
9. Emily M. Bender and Alexander Koller, “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data,” in Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (2020), 5185–5198. Why the reference never transfers gets the full argument in a companion essay, Borrowed Names: Why LLM Tokens Do Not Inherit Reference. The strongest current reply belongs to Dimitri Coelho Mollo and Raphaël Millière, “The Vector Grounding Problem,” arXiv:2304.01481 (2023), who distinguish five notions of grounding and argue that reinforcement learning from human feedback may supply the referential kind; the disagreement turns on whether feedback-shaped functions are world-involving in the right way, which is denied here on stake grounds: feedback shapes the model’s dispositions to the trainers’ satisfaction, so the operative norms belong to the trainers, and nothing in the exchange becomes right or wrong for the system itself. ↩
10. Gurnee et al., “Verbalizable Representations,” Outlook. The authors offer the convergence as evidence that workspace architecture reflects deep computational pressures rather than biological accident; the reverse reading given here and in the reply to the checklist objection below — attainable under text-only pressure, therefore no maker of meaning, and evidentially spent for a system of known provenance — belongs to this essay, not to them. ↩
11. Gurnee et al., “Verbalizable Representations,” section 9.3 (“Notable differences from human cognition”). The same section draws a hedged analogy to psychedelic ego-dissolution and meditative selfless states as human cases in which something continues to function without a foregrounded self, while noting that the base model offers a stable, directly inspectable instance of the dissociation rather than a transient, retrospectively-reported one. ↩
12. Ned Block, “On a Confusion about a Function of Consciousness,” Behavioral and Brain Sciences 18, no. 2 (1995): 227–287; “Consciousness, Accessibility, and the Mesh Between Psychology and Neuroscience,” Behavioral and Brain Sciences 30 (2007): 481–548. Block’s overflow argument, taken up on its own terms elsewhere (resisting the anti-representationalist conclusion he draws from it) in a companion essay, The Phenomenal/Access Distinction: Two Roles, Not Two Kinds; the present essay needs only the access/phenomenal distinction itself, not that further dispute. The absent-minded driver who opens this essay is the literature’s own stock case — David Armstrong’s long-distance truck driver, whose missing introspective awareness Fred Dretske dissects in Naturalizing the Mind (Cambridge, MA: MIT Press, 1995) — pressed into service here for the access/automatic contrast rather than for Armstrong’s higher-order moral. ↩
13. The stake and its consequences for machine intentionality get the full argument in a companion essay, Dennett and the Missing Stake. Compressed: a system has original, non-derived aboutness only where something can go wrong for the system itself, at its own cost, rather than merely for an external interpreter. Digital computation’s defining virtue — that a running process is a type restorable from its weights, never an irreplaceable token — is also the precise engineering-out of that cost. Nothing here rules silicon out in principle; it rules out substrate-indifference, which is a different thing. ↩
14. Patrick Butlin, Robert Long, Eric Elmoznino, Yoshua Bengio, Jonathan Birch, Axel Constant, George Deane, Stephen M. Fleming, Chris Frith, Xu Ji, Ryota Kanai, Colin Klein, Grace Lindsay, Matthias Michel, Liad Mudrik, Megan A. K. Peters, Eric Schwitzgebel, Jonathan Simon, and Rufin VanRullen, “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness,” arXiv:2308.08708 (2023). The report itself stops short of claiming that satisfying its indicators would settle the question outright; the inferential slide criticized here belongs to how the framework tends to get used, not to a claim its authors make in so many words. ↩

Get new essays by email