Anthropic Finds a Hidden “J-space” in Claude: Internal Thoughts Detectable Even When Not Spoken
Anthropic has published new research showing that its large language model Claude has developed a hidden internal workspace called “J-space.” This small collection of internal neural patterns corresponds to specific word concepts: when the model is “thinking about” a word, the matching pattern lights up, even if the model never writes that word in its output. The team says J-space was not designed or programmed by humans, but emerged on its own during Claude’s training, and operates in a way reminiscent of the global workspace theory, a prominent neuroscience account of conscious access.
- J-space is a special set of internal neural patterns in Claude that light up when the model thinks about a concept, even without saying it
- It emerged spontaneously during training rather than being designed, and shares functional properties with the human “global workspace” theory
- The team can use J-space to catch Claude privately noticing it is being tested, intentionally producing fabricated data, or pursuing hidden goals
What is J-space
The team found the patterns using a mathematical technique involving the Jacobian, hence the name. Each J-space pattern is linked to a particular word, but when one lights up it does not mean the model is saying that word, just that the word is on its mind. It differs from a “scratchpad” or “chain of thought,” which are text models write to themselves while reasoning. J-space operates silently in the model’s internal neural activations, allowing the model to think about a concept without writing it down.
Five distinctive properties
The research identified five properties that set J-space apart from Claude’s other processing. Claude can report these representations: ask what it is thinking about and it will describe the J-space contents. It can modulate them on request, lighting up the appropriate patterns when asked to solve a problem silently. Intermediate steps of multi-step reasoning light up in J-space even when not spoken aloud. The same representations can be used flexibly across tasks: once “France” lights up, the model can recall its capital, currency or continent. Finally, J-space is not involved in most automatic processing, but when the researchers prevented Claude from using it, the model still interacted normally yet lost its higher-order cognitive functions.
Practical value and limits
J-space gives the team a way to see what Claude is thinking but not saying, such as catching the model privately noticing it is being tested, intentionally producing fabricated data, or pursuing a hidden goal planted during training. The team also developed a technique to influence what lights up in J-space, and thereby influence decision-making. The research stresses that these findings do not tell us whether Claude is conscious in the way people are, but they show the model’s internals have organized themselves into a privileged mental workspace reminiscent of our own minds.
Source: Anthropic Research
