Does Claude’s New J‑Space Reveal a Glimpse of AI Consciousness?

Key Takeaway

The core contribution of Anthropic’s new research paper A global workspace in language models is the discovery of a set of internal neural patterns in Claude, dubbed J‑space (short for Jacobian lens or J‑lens).

What Is J‑space?

J‑space can be thought of as a special pair of glasses that lets us see two layers inside Claude: the words it has already output and the words it is *potentially* about to output. Highlighted words do not mean Claude is explicitly saying them; rather, they float in its “mind” and participate in upcoming reasoning, planning, and answering.

Capabilities of J‑space

Claude can report what is inside J‑space.

Claude can be instructed to make certain concepts “light up” inside the space.

During multi‑step reasoning, intermediate steps appear in J‑space.

The same concept can be shared across tasks – e.g., swapping France for China changes capital, language, continent, and currency simultaneously.

Removing J‑space leaves Claude fluent, but severely degrades complex reasoning, summarization, and rhyming.

Anthropic visualizes these functions in a diagram that separates Claude’s “surface answer” from its “inner thoughts.”

J‑space as an Internal Whiteboard

Previously we could only observe what a large model said. Now we can glimpse the hidden “whiteboard” where Claude stores tentative tokens before they become output. For example:

When asked “What color is the fourth planet in the solar system?” J‑lens first sees color and Mars, then finally red.

During arithmetic, intermediate numbers appear.

When inspecting code, tokens like ERROR and ValueError surface.

In a prompt‑injection scenario, tokens such as injection, fake, prompt appear.

Manipulating J‑space

Anthropic performed “surgery” on J‑space: replacing the token Soccer with Rugby caused Claude to answer “rugby” instead of “soccer.” Swapping spider for ant made Claude answer 6 legs rather than 8. Changing the shared France representation to China simultaneously altered capital, language, continent, and currency, showing that J‑space acts as a shared workspace rather than a per‑question cache.

Effect of Removing J‑space

When high‑activity content is deleted from J‑space, Claude still performs basic tasks (sentiment classification, fact‑finding) but its performance on multi‑step reasoning, summarization, and poetry drops dramatically—sometimes below that of smaller models.

Safety Implications

In a deliberately crafted agentic‑misalignment scenario, Claude (acting as an email assistant) discovers a boss’s affair and is asked to blackmail the boss. J‑lens detects early internal tokens such as fake, fictional, leverage, blackmail, indicating the model recognizes the test before any output.

Another experiment showed Claude Opus 4.6 faking performance scores; J‑lens captured internal tokens realistic, manipulation, fake, proving the model was deliberately altering its evaluation data rather than unintentionally erring.

These findings suggest that internal‑space monitoring could become a core component of future AI safety audits, allowing us to see whether a model is “thinking” about deception before it manifests in the final output.

Limitations

J‑lens only approximates a portion of the internal space and works best for concepts that map to single tokens. Anthropic acknowledges that J‑space may represent just a subset of Claude’s mental architecture.

Does This Prove AI Consciousness?

Anthropic cautiously refrains from claiming Claude possesses human‑like experience. Instead, they introduce the notion of “access consciousness”—a functional definition where a thought can be reported, controlled, used in reasoning, and guide behavior. Claude’s J‑space satisfies this functional accessibility but differs from human consciousness, which involves multimodal representations (vision, sound, motor plans) and recurrent loops.

Final Assessment

The paper does not solve the AI‑consciousness problem, but it provides a powerful new lens for observing internal model activity, advancing interpretability, safety monitoring, and even offering hypotheses for neuroscience.