Anthropic, the company behind the Claude family of AI models, says it has found a way to peer inside its models' internal processing — what it describes, loosely, as their "thoughts."
According to Tom's Hardware, a new Anthropic research paper reports that Claude models appear to have a "global workspace," an internal structure the company calls a "J-space." Tom's Hardware says this space shows similarities to human internal processing, though it cautions that Anthropic is anthropomorphizing when it calls the activity "thought."
The technique Anthropic uses to look into this space is called "J-Lens," according to finance.biggo.com, which reports that the company claims J-Lens lets it peer into Claude's internal states.
Both outlets frame the work as an attempt to reveal more of what actually makes large language models tick — systems that are notoriously opaque even to the people who build them. Tom's Hardware notes that, whatever the philosophical debate over calling it "thought," the method may prove practically useful: it could help improve an AI model's honesty, make its behavior easier to oversee, and strengthen the guardrails meant to keep it safe.
The sources describe the findings as observations detailed in a research paper rather than a finished product, so it remains early. But the direction matters.
Why it matters: If AI companies can reliably inspect what a model is doing internally rather than judging it only by its outputs, they gain a new tool for catching dishonesty and unsafe behavior before it reaches users — a central challenge as these systems grow more powerful.