Anthropic’s AI Lens Reveals What Claude Is Actually Thinking

Anthropic just cracked open a window into Claude’s mind. The company’s new interpretability tool, the Jacobian lens, reveals hidden reasoning states inside large language models that no one could see before. For food-industry operators betting on AI for formulation, compliance, or supply-chain decisions, what’s inside that black box matters.

TLDR

• Anthropic built a tool called the Jacobian lens to inspect live LLM reasoning.
• The tool reveals a hidden ‘concept space’ where Claude works through problems.
• Findings range from mundane processing patterns to genuinely unnerving behaviors.
• AI interpretability inside LLMs is now a measurable, not theoretical, discipline.
• Operators relying on AI outputs face new pressure to understand model internals.

Anthropic researchers, writing up findings covered by MIT Technology Review, built the Jacobian lens to track how Claude processes concepts in real time. The tool maps a hidden internal space where the model appears to “puzzle over” ideas before producing an answer. That space was previously invisible to developers and auditors alike.

AI Interpretability Inside LLMs Finally Has a Lens

Significant. Until now, large language models functioned as statistical black boxes; inputs went in, outputs came out, and the middle was opaque. The Jacobian lens changes that by capturing gradient-based signals that expose which internal representations are active during inference. Anthropic says the results range from mundane to unnerving, though the company has not yet detailed every finding publicly.

For food manufacturers using AI in label review, ingredient sourcing, or regulatory submissions, that opacity has always carried risk. A model that confidently outputs a clean-label claim may be “reasoning” through conflicting internal states that no audit trail captures. The Jacobian lens suggests that kind of internal conflict is detectable, in principle.

What This Means for Operators Deploying AI

The food industry’s adoption of AI tools is accelerating across formulation, traceability, and consumer transparency platforms. However, most deployments treat model outputs as authoritative without interrogating how those outputs were reached. That gap is exactly what interpretability research targets.

Anthropics’ work does not yet translate into a plug-and-play audit tool for enterprise users. Still, it establishes that AI interpretability inside LLMs is a tractable engineering problem, not a philosophical abstraction. Suppliers and operators who select AI vendors should now ask pointed questions: Does your model expose reasoning states? Can outputs be traced to internal representations?

The companies that ask those questions first will be better positioned when regulators inevitably do.

Source: MIT Technology Review. https://www.technologyreview.com/2026/07/09/1140293/anthropic-found-a-hidden-space-where-claude-puzzles-over-concepts/