Anthropic found a small set of internal representations that Claude can report, reuse, and reason with. Its new Jacobian lens estimates which concepts an activation could make the model say later. The resulting “J-space” appeared across Claude Sonnet 4.5, Haiku 4.5, and Opus 4.5, with some analyses on Opus 4.6. The causal tests are stronger than a visual pattern. Replacing “spider” with “ant” inside J-space changed a downstream answer from eight legs to six. Heavy ablation, which removed J-space activity, cut accuracy on a controlled multi-step task from 98% to 6%, while 65% of ordinary next-word predictions still matched the untouched model.
J-space accounts for less than 10% of activation variance, yet some parts of the network connect to its directions about 100 times more strongly than ordinary directions. That makes it a plausible shared working area rather than a transcript of everything Claude computes. It could help investigators spot hidden goals, prompt-injection recognition, or an intermediate mistake before it reaches the answer. Google DeepMind’s Neel Nanda independently replicated the core result on Qwen 3.6 27B, though he calls the lens useful but limited.
This is evidence for “access consciousness,” a functional ability to make information available for reasoning and report. It is not evidence that Claude feels anything. The outside commentaries stress that transformers lack the brain’s body, enduring memory, and recurrent activity. Anthropic’s released code also reads single-token concepts only, so this is an incomplete view of the model’s internal work.
Sources:
- Anthropic research summary
- Full technical paper
- Independent expert commentaries
- Jacobian lens code
- Interactive J-lens examples
Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service
Reuse
Citation
@misc{kabui2026,
author = {{Kabui, Charles}},
title = {Claude {Has} an {Internal} {Workspace,} {Not} {Proof} of
{Consciousness}},
date = {2026-07-19},
url = {https://toknow.ai/posts/claude-internal-workspace-not-consciousness/},
langid = {en-GB}
}