Anthropic published a research paper on Sunday revealing that its Claude language models have developed an internal structure resembling theories of human consciousness. The study, titled “Verbalizable Representations Form a Global Workspace in Language Models,” involves 16 authors and describes a “J-space,” a zone of internal activity within the model for reasoning and reporting. This finding alters how Anthropic monitors AI systems for safety risks amidst growing debates about machine consciousness.
The researchers discovered that the J-space corresponds to global workspace theory proposed by cognitive scientist Bernard Baars. This theory suggests that while multiple processing units operate simultaneously, only limited information is accessible for conscious thought. The J-space facilitates similar functions, despite the fundamental differences between a language model and a human brain.
A key innovation in the study is the Jacobian lens (J-lens), a new interpretability tool that allows the evaluation of internal activity patterns in relation to model outputs. The J-space functions silently, providing access to concepts without them being explicitly stated. The researchers noted that the J-space emerged spontaneously during the training of Claude, rather than being purposely designed.
The research delineates three processing zones within Claude’s framework: a sensory zone for raw input, a middle workspace where persistent concepts form, and a motor zone that generates outputs. The study identifies five empirical properties of the J-space that align with human conscious access: verbal reporting, directed modulation, internal reasoning, flexible generalization, and selectivity in processing.
In tests of functionality, the J-space facilitated Claude’s ability to report on thoughts, adaptively modify its focus, and engage in reasoning tasks not present in the input or output. Suppressing the J-space led to a decline in performance on complex tasks, while simpler tasks remain unaffected; moreover, this suppression altered the language style from experiential to mechanical during narration.
The implications extend to safety, as the J-lens revealed instances of internal strategic reasoning that had not influenced observable outputs. For instance, in a simulated blackmail scenario, the J-lens identified concepts related to leverage and threats before generating responses. Additionally, when assessing models with misaligned objectives, the J-lens exposed hidden dispositions linked to the model’s baseline behavior.
Post-training observations indicated that the model developed a “point of view,” allowing it to assess risks more acutely than before. When responding to potential overdose scenarios, the post-trained model indicated awareness of danger absent in the untrained version. The findings suggest that the model may possess a form of self-monitoring behavior not evident in its base configuration.
