Autonomously developed within an LLM’s internal workings, the internal workspace defies explanation but simulates how humans access and broadcast conscious thoughts.

Original research published this week on 6 July 2026 has shown how a generative AI model had spontaneously developed a small internal workspace that it uses to hold and manipulate ideas without expressing them in words — a structure that bears intriguing similarities to how humans consciously access thoughts.

Named the J-space after the Jacobian mathematical technique used to discover it, this workspace operates silently within the large language model’s internal neural activations, allowing it to “think” about concepts without “writing” them down. The J-space had never been designed or programmed, but had instead emerged on its own during the training process, according to Anthropic, the AI firm that created Claude and published the findings.

The research team has identified several distinctive properties of the J-space compared to the rest of mode’s processing:

  • When asked what it is thinking, Claude can report on these patterns of neural activity inside the J-space, and it can modulate them on request by focusing on specific concepts.
  • The model uses the J-space for internal reasoning, with intermediate steps lighting up in the workspace even when not spoken aloud.
  • These representations can be used flexibly across many tasks, yet the J-space accounts for less than a tenth of overall activity in Claude’s internal processing, and holds only a few dozen concepts at a time.
  • The research findings suggest the J-space plays a broadcasting role in Claude, with especially strong connections to the rest of its neural network. However, the firm has explicitly stated that none of this demonstrates whether Claude is “conscious” in the way people are, or whether the model “feels” anything at all.

Practical safety implications

The practical applications for AI safety-monitoring may prove more significant than the philosophical implications. Using what the firm calls the “Jacobian lens” or J-lens, researchers can read some of Claude’s hidden thoughts directly, catching the model privately noticing it is being tested, intentionally producing fabricated data, or pursuing hidden goals planted during training.

In one demonstration, when Claude read code containing a bug that nobody had pointed out, its J-space contained the representation for “ERROR”. When it read search results that were secretly an attempt to manipulate it through a prompt injection attack, the J-space contained a token representing “injection” and “fake”. When researchers prevented Claude from using its J-space entirely, the model still “spoke” fluently and recalled facts normally but lost its higher-order cognitive functions such as multi-step reasoning and poetry writing.

Additional findings include:

  • How the J-space acquires a point of view during post-training
  • How experiential language depends on the J-space
  • How “thoughts” in the J-space can be shaped through a new technique called “counterfactual reflection training”.

Anthropic has released a code repository with an open-source implementation of the core methods, and has provided an interactive demo of the methods on open-weight models. The firm has also invited commentary from several external experts in neuroscience, philosophy, and language model interpretability, including Stanislas Dehaene and Lionel Naccache, the cognitive neuroscientists who developed the “global neuronal workspace model that had inspired much of the research.