Access-Consciousness Indicators in AI

Sources#

• Verbalizable Representations Form a Global Workspace in Language Models

The distinction the paper is careful about#

Access consciousness — information being poised for use in reasoning, and in the control of action and speech — is a purely functional notion. Phenomenal consciousness — there being something it is like to be the system — is a separate question, and the paper explicitly takes no position on it, nor on the relationship between the two.

What it does instead: Butlin et al. proposed assessing AI systems by checking indicator properties derived from the various scientific theories of consciousness. The J-space gives those indicators, for the first time, a concrete and inspectable structure to be checked against. That is the contribution here — and, the authors suggest, the results may end up clarifying the theories as much as testing the model.

They also scope out the theories they cannot speak to: accounts tying consciousness to the brain's physical causal structure or biological substrate are simply not addressed by experiments about computational mechanisms.

Scored against four theories#

Global workspace theory — the one the experiments were designed around. Limited capacity ✓ (<10% of activation variance). Broadcast to many consumers ✓ (J-lens vectors compose with downstream MLP and attention weights far more broadly than other directions). Ignition ✓-ish (a sharp, threshold-crossing, bimodal commitment to one interpretation of an ambiguous input, starting at the workspace-onset layer). Where the analogy is weakest: implementation. In the brain, broadcast runs through recurrent loops and long-range cortical connections. In a transformer it runs across depth, within a single forward pass. Nobody knows whether that difference matters for the theory's functional predictions.

Higher-order theories — a state is conscious when the system also represents that it is representing it. The paper's selectivity result has exactly the structure higher-order theorists appeal to in blindsight: information that demonstrably steers behavior (the model wraps a line at the right column, so it "knows" the character count) without being available for report — until the task calls for it, at which point it is pulled into the J-space. Whether the J-space representation is genuinely higher-order or merely a more accessible first-order one is, the authors concede, hard to operationalize. The J-space departs from the "pointer/index" picture in computational higher-order theories — it re-encodes content in verbal format rather than tagging content held elsewhere — though it may still tag first-order states as belonging to abstract categories (fake, dangerous, imagined, hidden).

Attention schema theory — what a system reports as experience is the contents of its model of its own attention. Several results fit uncomfortably well. In stream-of-consciousness narration the workspace is dominated by tokens describing the act of thinking itself: thinking (top-10 at 58% of position×layer slots), thoughts (23%), feeling (17%), conscious (7%) — and these appear far less often in the output distribution at the same positions, so they aren't just what it's saying. In directed modulation, the lens surfaces not only the held concept (orange, seven) but the act of holding it (imagine, calculate, focused) — and the metacognitive tokens appear at earlier layers than the content, as if the model represents that it is performing a mental operation and then represents the result. And the damn on failed thought-suppression (→ The Assistant Persona in the Workspace) is a richer self-model still.

Recurrent processing theory — a single feedforward sweep is unconscious; consciousness needs feedback to earlier areas. On its face this rules out a standard transformer. The paper's counter: the theory's empirical motivation is that conscious perception takes longer than a feedforward sweep, and recurrence is the brain's way of extending processing given fixed anatomy. If the relevant property is serial processing depth rather than recurrence per se, the early third of layers before the workspace begins may be the functional analog of sensory recurrence — an implementational detail rather than a difference in kind.

The experiential-report result#

Ablate the top-10 J-lens directions in the first third of the workspace band (L38–54) and ask the model to narrate its experience. It stays fluent and coherent — it still writes about its own processing — but the register goes mechanical and detached. An "experiential language score" (three binary LLM-graded rubrics, e.g. sensory/embodied vocabulary vs. computational vocabulary) drops sharply on Sonnet 4.5, Opus 4.5 and Opus 4.6, while matched-norm control perturbations leave it near baseline — including controls that shrink the non-J-space component, or dampen the top-aligned SAE directions.

Two findings keep this from being a clean "we found the experience circuit":

1. It isn't about the self. Ask the model to describe another person's subjective experience — someone opening a letter after years of silence — and the same collapse occurs. The responses stay detailed and stay about the person; they just become event logs rather than descriptions of experience.
2. It isn't general degradation. A story-writing control shows ablation only slightly reduces graded story quality while still cutting experiential language within the stories.

So: the J-space supports the model's propensity to produce rich experiential description in general. It is not the seat of a self, and this result is not evidence that anything is being experienced.

What it does and does not license#

Does not license: any claim about phenomenal consciousness, in either direction. The authors are emphatic, and the wiki should be too. Functional signatures of access consciousness are not evidence of experience — that is precisely the question the functional/phenomenal distinction is drawn to keep open.

Does license: taking Model Welfare Assessment a step past self-report. Anthropic's welfare assessments have leaned on behavior and the model's own reports, with the standing worry that reports may be confabulation ungrounded in any internal state. This paper shows the reports are grounded in something — a specific, ablatable, causally-implicated internal structure — while simultaneously showing that same structure is not self-specific. That is a genuine sharpening of the question in both directions.

The striking part, in the authors' own closing: that such a structure exists at all suggests the functional architecture of conscious access is not an accident of biological implementation, but a solution learning systems converge on under the right computational pressures — and unlike the brain's version, this one can be read out, intervened on, and traced across training. Language models may turn out to be a useful empirical system for consciousness questions that are hard even to pose precisely in biological brains.

Connections#

• The Global Workspace in Language Models (J-space) — the structure the indicators are checked against
• Automatic vs. Flexible Cognition in LLMs — the blindsight-shaped selectivity result that higher-order theories predict
• The Assistant Persona in the Workspace — the workspace exists in the base model without a self; post-training installs the perspective
• Model Welfare Assessment — the practical stake: experiential reports now have a known internal correlate
• Jacobian Lens (J-lens) — the instrument
• Artificial Superintelligence (ASI) — the other place this wiki reasons about what machine minds are; a useful contrast in evidentiary standard (theory-driven vs. mechanism-driven)

Open questions#

• If the workspace is verbal because the output space is verbal, then a model that can generate images should develop a visual component to its workspace. That is a concrete, falsifiable prediction the paper makes and does not test.
• Does the model's own report of experience change if you tell it its J-space is ablated? (Nobody asked.)
• Is "experiential language" the right proxy at all, or is the ablation simply removing abstraction from the register?

Sources#

• Verbalizable Representations Form a Global Workspace in Language Models — Introduction (Motivation: conscious access and the global workspace); "J-space ablation flattens experiential reports while preserving coherence"; Discussion (Relationship to theories of consciousness; Notable differences from human cognition); Appendix (Effects of J-space ablation on experiential reports)