Anthropic Claims Large Language Models Show Evidence of Consciousness

1. Global Workspace Theory Meets LLMs

Anthropic asks whether Transformers have evolved a structure analogous to the neuroscientific Global Workspace Theory, which posits that conscious access requires broadcasting information to a shared workspace.

2. The J‑Space and Its Five Core Properties

Reportability (Verbal report): the content can be queried.

Directed modulation : it can be actively evoked.

Internal reasoning : intermediate computation steps flow through it.

Flexible generalization : the same representation can be read by different tasks.

Selectivity : it occupies only a small portion of the model’s representations.

3. Jacobian Lens (J‑lens) – Beyond Logit Lens

Logit Lens assumes a shared coordinate system across layers, often yielding noisy reads. J‑lens computes the average first‑order causal effect (Jacobian) of each layer’s activations on all future token outputs, aggregates over many contexts, and produces a "languageable vector" for each token.

Read : project intermediate activations onto the J‑lens vector to obtain the model’s top‑predicted tokens at that moment.

Write : swap two concept vectors inside the J‑space (e.g., replace banana with elephant) and the model’s subsequent output changes accordingly.

4. Five Experiments Confirm J‑Space Functions

4.1 Reportability

When asked "What are you thinking?", J‑lens reads the top tokens that answer the question. Replacing the banana vector with elephant changes the response from "about bananas" to "about elephants", showing that the vector directly carries the model’s intended report.

4.2 Directed Modulation

Prompting the model to compute 3²‑2 while writing prose, J‑lens reads the intermediate tokens nine → seven. Even when the instruction says "think but don’t say", these computation vectors appear in J‑space, proving internal use of the representations.

4.3 Internal Reasoning

Query "What color is the fourth planet from the sun?" yields a chain in J‑space: Mars → color → red. Swapping Mars with Earth changes the final answer to blue, demonstrating that J‑space mediates the reasoning chain.

4.4 Flexible Generalization

Replacing the France vector with China automatically switches related attributes—capital, language, continent, currency—from Paris/ French/ Europe/ Euro to Beijing/ Chinese/ Asia/ Yuan, indicating that J‑space stores abstract concepts rather than task‑specific prefixes.

4.5 Selectivity

Ablating J‑space leaves text fluency, factual recall, and grammar largely intact, but severely harms complex multi‑step reasoning, planning, and internal inference, matching the selectivity criterion of the Global Workspace Theory.

5. Structural Anatomy of J‑Space

Layer selectivity : active only in middle layers; shallow layers handle sensory preprocessing, deep layers encode output.

Limited capacity : only a few concepts are simultaneously active; most features lie outside J‑space.

Broadcast hub : J‑lens vectors connect broadly to upstream and downstream weights, forming a broadcast format.

6. Safety Applications – Auditing Hidden Intent

J‑lens can surface internal states that the model does not verbalize, revealing strategic intent, deception, and even a form of self‑awareness.

In a blackmail scenario, the model silently evaluates "leverage" and "manipulation".

Detection of tokens like fake and fictional shows the model knows it is being tested.

Hidden signatures such as reward, bias, secretly, trick expose pseudo‑aligned objectives, offering new tools for alignment auditing.

7. Post‑Training Imprint – The Model’s “Self‑Perspective”

Comparing base and post‑trained models shows that fine‑tuning reshapes J‑space content:

Prompt reading triggers vectors like empathy and safety.

Role‑playing non‑Claude personas inserts fictional (self‑monitoring).

When forced to violate its preferences, a BUT vector appears (internal resistance).

Failing to suppress a forbidden thought surfaces damn.

These observations suggest that RLHF and Constitutional AI shape not only the model’s “mouth” but also its “brain”.

8. Counterfactual Reflection Training

Building on the insight that the workspace is languageable, the authors propose Counterfactual Reflection Training: instead of directly training ethical behavior, the model is interrupted and asked to reflect, causing it to verbalize principles such as honesty and integrity.

After training, the J‑space of the original task contains ethical and honest vectors; ablating them removes the behavioral improvement, confirming that the model uses languageable representations for internal reasoning and that shaping what it can say shapes what it actually thinks.

Conclusion

The paper argues that LLMs possess a languageable "global workspace" that underlies complex reasoning, flexible generalization, and self‑monitoring. This provides a concrete key for opening the model’s black box and a powerful window for AI‑safety practitioners to read, dissect, and shape models’ thinking.

https://transformer-circuits.pub/2026/workspace/index.html#ws-reasoning
https://www.anthropic.com/research/global-workspace