We just figured out how AI actually works (J-Space)

2026-07-08 · Source: Matthew Berman · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Intermediate, extended

Summary

Anthropic has unveiled the "J-space," an emergent internal global workspace within large language models like Claude, where the model's true internal reasoning and "conscious-like" thoughts occur, distinct from its direct output. This J-space, which was not explicitly programmed, accounts for less than a tenth of Claude's internal processing but is crucial for higher-order cognitive tasks such as multi-step reasoning, summarization, and poetry generation. Experiments demonstrate that the J-space is reportable, modifiable, and causally influences the model's final answers. For instance, changing a thought in the J-space from "soccer" to "rugby" directly altered Claude's reported thought. This discovery offers unprecedented interpretability into model behavior, revealing hidden thoughts and intentions, and has significant implications for AI alignment and control.

Key takeaway

For AI scientists and machine learning engineers focused on model interpretability and alignment, understanding Anthropic's J-space is critical. This internal workspace provides a direct window into a model's "true thinking," enabling you to detect hidden intentions, evaluate misalignment risks, and potentially influence model behavior through targeted training. You should explore methods to integrate J-space analysis into your safety evaluations, particularly for complex, high-stakes AI applications, to ensure models behave as intended even when unobserved.

Key insights

The J-space is an emergent internal workspace in LLMs where true, conscious-like reasoning and thoughts occur.

Principles

• LLM internal reasoning is distinct from output.
• J-space is causally linked to model behavior.
• Higher-order tasks rely on J-space processing.

Method

Anthropic's research involved probing and surgically modifying the J-space representations within Claude's neural network to observe causal effects on its internal thoughts and external responses.

In practice

• Monitor J-space for hidden model intentions.
• Influence model behavior via J-space training.
• Analyze J-space for multi-step reasoning traces.

Topics

• Anthropic
• J-space
• Language Models
• AI Alignment
• Model Interpretability
• Cognitive Architectures

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Ethicist

Related on AIssential

• Anthropic's new "J-lens" reveals a silent workspace inside Claude that mirrors a leading theory of consciousness — Claude models spontaneously developed an internal "J-space" mirroring human global workspace theory, detectable via a new "J-lens" interpretability tool.
• Is J-Space the Breakthrough AI Alignment Has Been Waiting For? — J-Space offers a causal window into language models' internal reasoning, enabling detection and influence of hidden cognitive states.
• So AI is "thinking" for real — Anthropic's J-space reveals AI's internal "thoughts" for complex reasoning, distinct from its direct output.
• No Space Like J-Space — The J-space is a functional global workspace in LLMs, containing verbalizable representations crucial for internal reasoning and output determination.
• Anthropic's New J-Space: Debunked? — Anthropic's "J-Space" is a computational routing mechanism in LLMs, not a conscious workspace or new mathematical subspace.
• 😼 Anthropic found Claude’s hidden workspace — Anthropic discovered Claude's "J-space," an internal neural workspace for silent reasoning and concept manipulation.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Matthew Berman.