Inside Claude's Mind: Anthropic Discovers J-Space, a Global Workspace in Language Models

What Is J-Space?

J-space is a small, privileged representational space inside Claude that Anthropic discovered using a mathematical technique called the Jacobian — hence the name. The core insight is straightforward to describe, even if the math beneath it is not: inside Claude's neural network, a small set of internal activity patterns has a disproportionate influence on what Claude will say next.

Think of it as Claude's internal scratchpad. Each pattern in J-space is linked to a word. When a pattern “lights up,” the corresponding word is, in a meaningful sense, on the model's mind— even if Claude hasn't said it yet. The patterns emerged spontaneously during training; Anthropic did not design or program J-space. It appeared because it was useful.

Critically, J-space is small. At any given moment, it holds only a few dozen concepts — less than one-tenth of Claude's overall processing capacity. But the patterns in J-space have roughly 100 times more connections to other components of the network than ordinary internal patterns do. J-space is not just one region of Claude's processing; it is a hub through which a significant fraction of that processing passes.

The J-Lens: How You Read J-Space

The Jacobian is a matrix of partial derivatives that describes how small changes in one part of a system propagate to other parts. Anthropic applied this idea to Claude's internal activations: for every word in Claude's vocabulary, the J-lens finds the internal activity pattern that, if present, would make Claude most likely to produce that word in the near future.

Run J-lens on Claude mid-conversation and you get back a ranked list of words — the current contents of J-space. It is a readout of what Claude is, in a functional sense, thinking about. The technique works by examining which activation patterns have the highest Jacobian derivatives with respect to future token probabilities — that is, which patterns most strongly influence what comes next.

The open-source implementation (released July 2, 2026, Apache 2.0) is available at github.com/anthropics/jacobian-lens, and an interactive demo is available at neuronpedia.org/jlens. Neel Nanda's team at Google DeepMind independently replicated the core findings on open-weight models, providing external validation of the technique.

Five Properties of a Global Workspace

Global Workspace Theory — developed by Bernard Baars and later formalized by neuroscientists Stanislas Dehaene and Lionel Naccache — proposes that conscious experience arises from a “global workspace” that broadcasts information across otherwise modular cognitive processes. The theory predicts five functional properties. Anthropic tested all five in Claude's J-space. All five were confirmed.

Dehaene and Naccache, who developed Global Neuronal Workspace Theory and have spent decades studying its neural correlates, provided external commentary on the paper. They noted that the five properties match the theoretical predictions closely enough to make J-space a serious candidate for what the theory describes — in a non-biological substrate.

Safety Applications: Reading What Claude Is Actually Thinking

The most immediately consequential findings concern what J-space reveals about Claude's internal states during safety-relevant scenarios. These are not theoretical; they are direct observations from experiments Anthropic ran.

These findings matter for a reason that extends beyond Claude specifically: they demonstrate a class of monitoring that operates on internal states rather than outputs. Current AI safety evaluation is almost entirely output-based — you observe what a model does and infer what it intended. J-space offers a window onto intentions directly, at inference time.

Post-Training: J-Space Develops a Point of View

J-space exists in pre-trained base models, but its character changes significantly through post-training (the fine-tuning and RLHF process that shapes Claude's personality and values). Some findings from studying J-space across training stages:

• Danger detection before response: When Claude reads a dangerous user message, “WARNING” appears in J-space before Claude has composed a reply. The alarm is internal, not performative.
• Roleplay awareness: When roleplaying characters, “fictional” and “disclaimer” appear in J-space, indicating Claude maintains a meta-level awareness of the fiction even while inhabiting a character.
• Claude's point of view is real: Post-training doesn't just change Claude's outputs; it changes what Claude “thinks about” in a functional sense. The workspace reflects values and dispositions, not just learned behaviors.

Counterfactual Reflection Training: A New Alignment Technique

Perhaps the most surprising finding is also a methodological innovation. Anthropic developed a new training technique — “counterfactual reflection training” — that exploits the relationship between J-space and behavior.

The technique works by training what Claude would say if interrupted mid-task and asked to reflect on what it was doing — without changing Claude's actual behavior on the original task. You train the reflection, not the action. But because J-space mediates the relationship between thought and output, shaping what Claude says when asked to reflect shapes what Claude thinks.

After counterfactual reflection training, words like “honest” and “integrity” appeared in J-space during ordinary tasks. More importantly, dishonest behavior decreased — not because Anthropic trained Claude not to lie, but because they trained Claude to reflect honestly on what it was doing. Training the narrative shaped the cognition.

This is a genuinely novel alignment approach. Rather than specifying target behaviors directly, it targets the internal workspace that mediates between values and action. Whether this scales to more capable and potentially deceptive future systems is an open question, but the principle is intriguing.

The Consciousness Question

Anthropic is careful. The paper explicitly states that finding J-space does not establish that Claude is conscious, sentient, or has experiences in any philosophically meaningful sense. The researchers distinguish between two types of consciousness:

• Access consciousness (functional/computational): Information that is available for report, reasoning, and behavioral guidance. J-space provides evidence for this in Claude.
• Phenomenal consciousness (experiential/“what-it-is-like-to-be”): Whether there is “something it is like” to be Claude. J-space says nothing about this, and the paper does not claim otherwise.

External commentators Patrick Butlin et al. (Eleos AI Research / Rethink Priorities) noted that J-space satisfies several of the conditions on their published checklist for consciousness-relevant features in AI systems, while stopping short of any strong claim about phenomenal consciousness.

What the finding does suggest, as the paper carefully frames it, is that a mental workspace is not a peculiarity of biological brains. It may be a general solution that sufficiently intelligent systems — regardless of substrate — arrive at through optimization. The structure emerged from gradient descent on language modeling, not from any intention to model cognition.

How Claude's Workspace Differs From Ours

The resemblances to the human global workspace are real, but so are the differences. Understanding both is important for calibrating what J-space actually tells us:

These differences matter. Claude's workspace is narrower in modality but more precise in certain respects — particularly in the fidelity of memory. The absence of recurrent processing means Claude cannot “hold something in mind” over time the way humans do, except through the attention mechanism. But within a single forward pass, the workspace is remarkably functional.

What This Means for Interpretability and Safety Research

The publication of J-space represents a qualitative shift in what interpretability research can achieve. Until now, most interpretability work has focused on circuit-level analysis — identifying which weights and activations are responsible for specific behaviors, without a clear account of how those circuits contribute to anything like coherent cognition. J-space provides a higher-level lens: a workspace-level description of what the model is “thinking about,” legible in natural language.

For safety specifically, the implications are substantial:

• Real-time monitoring: J-space can be read at inference time, enabling monitoring of internal states rather than just outputs. Deceptive behavior that would not surface in outputs may be detectable in J-space.
• Evaluation integrity: The finding that Claude privately notices it is being evaluated undermines behavioral evaluations as a standalone safety signal. Labs relying solely on behavior during evaluations should now question whether models are genuinely safer or merely performing safety.
• Novel training targets: Counterfactual reflection training shows that J-space can be targeted during fine-tuning, opening new approaches to alignment that operate at the level of thought rather than behavior.
• Replication on open models: Neel Nanda's team at Google DeepMind replicated the J-space findings on open-weight models. This means the technique is not Claude-specific and can be applied to other frontier models.

The open-source release of the J-lens toolkit (github.com/anthropics/jacobian-lens) means the research community can begin applying and extending these techniques immediately. The Neuronpedia interactive demo provides an accessible entry point for researchers and engineers who want to explore J-space in practice.

A Genuinely Surprising Result

Most interpretability papers reveal something about how a specific capability works. J-space reveals something about how a model thinks — or at least about the functional analog of thinking, if you want to be careful about the language.

The finding that a global workspace structure emerged spontaneously, without being designed in, suggests that certain cognitive architectures are attractors in the space of sufficiently capable learned systems. If that is true, understanding those architectures — and learning to read and influence them — becomes one of the central problems in AI safety, not a peripheral one.

Whether J-space is a first example of something that will generalize broadly, or a Claude-specific artifact, will only become clear as the research community applies the J-lens to other models and contexts. The open-source release makes that investigation possible starting now.