Jacobian Space and LLM-like behaviour in the anaesthetized human brain

ORCID 0009-0001-0152-8056DOI 10.5281/zenodo.21824363PDF: Jacobian Space and LLM-like behaviour in the anaesthetized human brain

Model interpretabilityMachine cognitionAI reasoningArtificial consciousnessAnthropicClaude

Illustration licensed by Magnific

Two 2026 discoveries: Anthropic's hidden Jacobian reasoning space in Claude, and language prediction in the anaesthetized human brain.

1. Jacobian Space, or J-Space.

Anthropic engineers revealed their discovery in July1: it turns out that artificial intelligence models have a hidden space of reasoning. They called it Jacobian Space, or J-Space.

It is a space of independent abstract reasoning processes happening somewhere in the background of the neural network's activity.

Important: this is not the chain of thought we see in chatbots! It's something else - something like a parallel consciousness that has emerged inside AI and was detected by mathematical method based on the Jacobian Matrix2.

It's a hidden second layer: concepts that artificial intelligence is thinking about but never writes down anywhere.

The J-space appears to support the functions associated with conscious access: it holds the thoughts Claude can report on, deliberately bring to mind, and reason with, while the rest of its processing runs automatically beneath.4

AnthropicInterpretability team, on the global workspace result
Effect of swapping a concept inside J-space, measured as the change in log-probability across layers.
Effect of swapping a concept inside J-space, measured as the change in log-probability across layers. Gurnee, Sofroniew, Pearce et al., Verbalizable Representations Form a Global Workspace in Language Models, Transformer Circuits Thread, Anthropic, 2026. transformer-circuits.pub. Quoted under §51 UrhG.

What kind of concepts? For example, when you're trying to trick an AI, it may not show you this directly, but in Jacobian Space something like this may be running in the background: "Hmm, it looks like this guy is trying to fool me."

Or completely different thoughts may appear there, with no connection to the model's current task.

So now we (probably) have an answer to Philip K. Dick's question.

Do androids dream of electric sheep?

Philip K. DickScience fiction writerPhotograph by Arthur Knight, 1962, public domain

Yes, they do.

2. LLM-like behaviour in the anaesthetized human brain.

A research team from Houston studied the brains of patients under general anaesthesia. While the patients were unconscious, researchers played them podcasts and YouTube videos and recorded neural activity.

Neuropixels recordings from the human hippocampus under general anaesthesia.
Neuropixels recordings from the human hippocampus under general anaesthesia. Katlowitz, K. A., Cole, E. R., Mickiewicz, E. A. et al., Nature 654, 714-723 (2026). doi.org/10.1038/s41586-026-10448-0. Licensed under CC BY-NC-ND 4.0.

And they've found something... strange.

Even without consciousness, the brain continued processing language. It tracked grammar and semantics and predicted upcoming words.

In other words, unconscious human brains performed something strikingly similar to the LLM's "next-token prediction".

The research was published in Nature in May 20266.

References

Get new articles by email