AI ‘thinking’ words it never says: What this tells us about consciousness

Artificial intelligence (AI) debates often hinge on defining what "intelligence" truly means, leading to renewed discussion after Anthropic's research found the chatbot Claude Sonnet 4.5 exhibiting what appeared to be an internal thought process. This discovery reignited the debate with a book-length research paper and a blog post announcing the findings, titled "A Global Workspace in Language Models".

The research identified a collection of neural patterns within Claude's model, dubbed the "J-space" or Jacobian space, which play a special role in comparison to other internal processing. These patterns provide a mathematical approximation of the language model's working memory, representing particular words without the model explicitly stating them.

However, this does not necessarily reveal that Claude possesses consciousness. Anil K. Seth, a professor of cognitive and computational neuroscience, doubts that AI can achieve consciousness, while AI pioneer Geoffrey Hinton argues that chatbots possess subjective experiences. Some researchers suggest that bio-hybrid computers, which incorporate human neurons on silicon, may be more likely to achieve sentience.

Inside Claude's processing, when the model is asked to count and introspect, different words pop up in its internal activations, such as "countdown" and "done". Anthropic interprets this as evidence of internal thought processes, representing concepts that Claude eventually outputs but are not visible to the user. While this step is a promising first step in understanding conscious access in language models, further research is needed to understand how the J-space works.