cd /news/artificial-intelligence/anthropic-finds-hidden-thinking-word… · home topics artificial-intelligence article
[ARTICLE · art-121836] src=insideai.news ↗ pub= topic=artificial-intelligence verified=true sentiment=· neutral

Anthropic Finds Hidden ‘Thinking’ Words Inside Claude Sonnet 4.5

Anthropic researchers have discovered internal neural patterns in Claude Sonnet 4.5, named J-space, that act as a silent working memory, with hidden words like 'countdown,' 'halfway,' and 'done' appearing during a counting task without reaching user-facing output. The findings, published in a research paper and blog post titled 'A Global Workspace in Language Models,' reignite debates on machine consciousness, with experts like Anil K. Seth of the University of Sussex cautioning against claims of sentience while noting bio-hybrid computers from Cortical Labs may offer more promise.

read4 min views2 publishedSep 7, 2026
Anthropic Finds Hidden ‘Thinking’ Words Inside Claude Sonnet 4.5
Image: Insideai (auto-discovered)

September 7, 2026, (Inside AI) — Anthropic researchers have uncovered internal neural patterns in Claude Sonnet 4.5 that resemble a silent thought process. The patterns, named J-space, emerged during training without explicit programming.

The discovery centers on a subset of the model's inner computations that act like a mathematical working memory. When researchers asked Claude to count to five while introspecting, hidden words such as "countdown," "halfway," and "done" appeared in its neural layers. These words never reached the user-facing output.

This finding reignites a core question: can algorithms ever become conscious? The answer splits the scientific community.

Anil K. Seth, Professor of Cognitive and Computational Neuroscience at the University of Sussex, rejects any leap toward consciousness. He argues that biological traits may be necessary for subjective experience. Geoffrey Hinton, an AI pioneer, has repeatedly claimed chatbots already have subjective experience.

The J-space operates silently within neural activations. It lets the model represent a concept without writing it down. Anthropic stresses it was not designed or programmed. It emerged on its own during Claude's training.

Dave Bergmann, an Anthropic researcher, compares the process to cooking a meal while narrating only some steps aloud. The inner monologue stays hidden. The spoken steps are incomplete. In technical terms, the visible reasoning is token space. The hidden reasoning is J-space.

Hidden Words Reveal a Silent Workspace #

During the counting experiment, researchers parsed the model's many layers of artificial neural networks. These layers had long remained impenetrable. What surprised them was the appearance of words tied to the task but never spoken.

"Countdown" was one such word. Midway through the output, "halfway" appeared internally. After the model said five, the word "done" lit up in its neural layers. Anthropic interprets this as evidence of an internal thought process.

J-space is short for Jacobian space. It is named after the mathematical technique used to find the patterns. Each J-space pattern links to a particular word. When a pattern activates, the model is not saying that word. The word is simply represented in internal processing.

Anthropic published a book-length research paper on large language model internals. A blog post announced the findings in July, titled "A Global Workspace in Language Models." The company describes J-space as a small collection of internal neural patterns that play a special role compared to all other processing.

Biological Hybrids May Shift the Debate #

Seth points to a different path. Australian startup Cortical Labs is building bio-hybrid computers with human neurons mounted on silicon. He sees greater promise there for sentience.

"The kind of bio-hybrid computers that Cortical Labs and others are building may well move the needle on the potential for consciousness. The reason is that, all else equal, the more similar our technologies are to real brains, the more plausible consciousness becomes. The challenge is that we don't yet know what difference makes the difference," Seth told The Indian Express.

Anthropic is careful to avoid overclaiming. The J-space experiments do not show that Claude can have experiences or feel things like humans. There is still no evidence of consciousness. But J-space appears to support functions tied to conscious access.

It holds the thoughts Claude can report on, deliberately bring to mind, and reason with. The rest of its processing runs automatically beneath. Anthropic calls this a first step in an extensive line of research.

"This work is just a first step in what we expect to be an extensive line of research. The J-space looks like a good candidate for the divide between consciously accessible and unconscious processing in a language model, but we'd be surprised if it's the whole story... And there remain many mysteries about how the J-space works. We don't know what mechanism decides what enters the J-space in the first place. We've seen hints that it's tied to Claude's sense of self, something like emotional reactions, and traces of metacognition, without exactly having worked out how. But we now have methods for tackling questions like these. As that work progresses, our understanding of LLM minds -- and their relationship to our own -- will grow clearer," Anthropic said in its blog post.

The finding does not settle the consciousness debate. It sharpens it. The gap between hidden processing and reported experience remains wide. But the tools to probe that gap are now emerging.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @anthropic 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/anthropic-finds-hidd…] indexed:0 read:4min 2026-09-07 ·