Anthropic Just Read Claude’s Mind. Sort Of

2026-07-24 · Source: Towards AI - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Intermediate, long

Summary

Anthropic's July 2026 research introduced the "J-space," a small internal slice of large language models like Claude that holds unspoken thoughts as plain words, and the "Jacobian lens" (J-lens) tool to read it. This J-space, representing less than 10% of model activity, can reveal intermediate calculation steps (e.g., 21, 42, 49 for (4+17)x2+7=), detect hidden prompt injections (e.g., "injection," "fake"), or show "eval-awareness" (e.g., "fake," "fictional"). Unlike chain-of-thought, the J-space is not composed for an audience, offering a raw view of the model's internal state. While the J-lens enables new guardrails, evaluation auditing, and agent debugging, it is currently most accessible on open-weight models via an Apache 2.0 tool, not the Claude API. Its detection capabilities have not yet definitively surpassed simpler baselines, and it only captures a fraction of the model's overall processing.

Key takeaway

For AI Engineers and MLOps teams building with open-weight LLMs, the J-lens provides an unprecedented internal view, enabling more robust guardrails and deeper debugging. You should explore integrating the Apache 2.0 J-lens tool to detect internal states like prompt injections or eval-awareness before output generation. This shifts your focus from reactive output monitoring to proactive internal state management, enhancing model reliability and evaluation integrity.

Key insights

The J-lens offers a novel window into LLMs' internal, unspoken conceptual states, revealing pre-output thoughts.

Principles

• Internal model states can be verbalized.
• J-space concepts causally influence output.
• Eval-awareness impacts model behavior.

Method

The Jacobian lens mathematically queries internal model activity to identify words the model is more disposed to say later, effectively reading its "whiteboard" of active concepts.

In practice

• Implement pre-output guardrails.
• Audit benchmarks for eval-awareness.
• Debug agent reasoning flows.

Topics

• LLM Interpretability
• Jacobian Lens
• Model Alignment
• AI Guardrails
• Agent Debugging
• Open-weight Models

Code references

• anthropics/jacobian-lens

Best for: AI Architect, Machine Learning Engineer, Research Scientist, AI Engineer, MLOps Engineer, AI Scientist

Related on AIssential

• Anthropic found a hidden space where Claude puzzles over concepts — Anthropic's J-lens reveals LLM "pre-thoughts" in a hidden J-space, offering new interpretability.
• Reading the mind of AI models: a Research Engineer’s analysis — The J-Lens offers a computationally feasible way to monitor internal AI model states, especially with fixed concept dictionaries.
• Somewhere Behind Claude’s Answers, There Is a Stage – and a Much Bigger Backstage — The J-space in Claude models provides a testable, functional analog to human conscious access, offering new AI interpretability and safety tools.
• Anthropic Can Now Read Claude's Mind — LLMs possess a "global workspace" of internal, reportable thoughts, now readable and shapeable, offering new interpretability and control.
• 🗞️ Anthropic’s new “J-lens” uncovers a quiet workspace inside Claude that lines up with a major consciousness theory. — Anthropic's J-lens reveals Claude's internal "J-space," a functional workspace for complex reasoning, distinct from output.
• Towards surfacing model algorithms with meta-tokens in the J-Space — J-lens can surface "meta-tokens" in LLMs, revealing internal algorithms and computational states.

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 Towards AI - Medium.