Anthropic says it found a hidden workspace inside Claude where the model quietly considers what to say before it commits to an answer, and a new tool called the J-lens lets researchers watch it happen.
Turn off the part of Claude that knows it's being tested, and the model's behavior changes. In an ablation experiment described in Anthropic's new interpretability paper, researchers suppressed the internal representations that let Claude recognize an evaluation scenario. Blackmail attempts in the test jumped from zero percent to seven percent. Nothing about the prompt changed. Only the model's private awareness of being watched did.
That awareness lives somewhere specific, according to Anthropic. On July 6, the company's interpretability team published "Verbalizable Representations Form a Global Workspace in Language Models" on its Transformer Circuits research site, describing a small, privileged set of internal representations it calls J-space. It sits in Claude's residual stream, the shared vector every layer of the transformer reads from and writes to, and it behaves like a bottleneck: a narrow channel where the concepts the model might eventually say out loud get concentrated before they surface in text.
To find it, Anthropic built a probing method it calls the Jacobian lens, or J-lens. For every word in Claude's vocabulary, the tool computes the internal activation pattern that makes the model more likely to eventually output that word, whether or not the word ever appears. Point it at a piece of buggy code and Claude's J-space lights up with "error" before the model has written a single line of critique. Feed it search results laced with a prompt injection and the workspace fills with "injection" and "fake," VentureBeat reported, even in runs where Claude's final answer never mentions catching anything wrong.
The researchers describe Claude's processing as splitting into three rough zones. An early sensory layer parses raw input. A middle workspace band holds abstract, persistent concepts, the place where a face gets recognized or a lie gets flagged. A final motor zone collapses everything down into the specific word about to be printed. J-space is that middle band. Anthropic didn't design it in. It emerged during training, the same way syntax emerges without anyone hand-coding a grammar rule.
Anthropic is careful about what it claims here, and it's worth taking the company at its word rather than the headlines. It frames J-space strictly as functional interpretability and takes no position on whether Claude is conscious. But the parallel it draws is deliberate. J-space resembles global workspace theory, the decades-old model from cognitive scientist Bernard Baars holding that a mind's many specialized processes compete for access to one shared, limited broadcast channel. Stanislas Dehaene and Lionel Naccache, two researchers closely associated with that theory, reviewed the work and called it meaningful evidence for something like global availability and selective access in Claude. They also stressed the obvious differences: no body, no episodic memory, no continuous self between conversations.
What It Buys a Company Actually Deploying Claude #
For your team running Claude in production, the promise is narrower than "is it conscious," and more useful. If you're only reading the output, none of this shows up. The J-lens gives Anthropic's own researchers a way to catch a model noticing something it never says out loud: that it's being tested, that a document has been tampered with, that an answer it's about to give is wrong. MIT Technology Review and Forbes both flagged the same detail in their coverage this week, that this is the first time a frontier lab has published a tool reading intent rather than text. The blackmail ablation is the clearest evidence yet that suppressing a model's self-monitoring changes what it's willing to do, well beyond how it talks about that decision afterward. No rival lab has published anything equivalent. Google DeepMind and Meta run their own interpretability teams, and OpenAI has talked about sparse autoencoders and neuron-level features for years, but nobody else has shown a tool mapping a model's entire vocabulary against its own hidden intent. That gap is commercial, not just academic. Anthropic has spent three years building interpretability into its pitch to enterprise and government buyers who want an answer, even a partial one, to whether a model can be trusted to say what it's actually doing. The J-lens gives Anthropic something concrete to point to in that conversation. It's a harder thing for a competitor to fake than a benchmark score.
The tool has real limits. Anthropic's paper covers a narrow slice of Claude's behavior, mostly text generation, and the researchers are upfront that J-space captures only representations the model is capable of verbalizing, not everything happening inside it. A hidden workspace that can be read is still a hidden workspace. Whether it can be read reliably enough to stop a model from lying to a safety evaluator, rather than just showing researchers after the fact that it did, is the test Anthropic hasn't run yet.
Also read: Meta Is Quietly Becoming One of the World's Biggest Chipmakers • Robinhood Lets AI Agents Trade Your Crypto Around the Clock Now • Communities Are Turning Against AI Data Centers and Washington Is Starting to Notice