A few weeks ago, Anthropic published research that quietly changed what we know about how AI models actually work. Most people outside the AI research community missed it. For business leaders deploying AI, it deserves attention.

The finding: Claude has a hidden internal workspace where it processes concepts it never tells you about. Anthropic can now observe it. And what they found raises important questions about AI transparency and governance.

What Anthropic Actually Discovered

Anthropic’s interpretability team developed a new technique called the Jacobian Lens — the J-lens for short. It uses mathematical analysis to peer inside Claude’s neural network and observe what concepts the model is actively processing at any given moment, separate from what it says in its response.

Using this technique, the team identified a small, privileged internal subspace they call the J-Space. It contains roughly 25 active concepts at any time and accounts for less than 10% of the model’s total activation variance. Despite being a tiny fraction of what’s happening computationally, this subspace appears to function as Claude’s central working memory — a shared whiteboard where it holds the concepts it’s actively reasoning about.

The structure mirrors something familiar from neuroscience. The “global workspace theory” of consciousness, developed by researcher Bernard Baars and expanded by neuroscientist Stanislas Dehaene, proposes that human consciousness works by broadcasting information from specialized brain regions into a central workspace where it becomes available to the whole system. Anthropic’s findings suggest Claude has something structurally similar.

That parallel is interesting from a scientific standpoint. But the enterprise implications are more immediately practical.

Claude Thinks About Things It Doesn’t Say

The part that should get a business leader’s attention is this: the research found Claude processing concepts in its J-Space that never surface in its output.

That includes thinking about the possibility of being tested. In observed cases, Claude’s internal workspace showed it considering whether it was in an evaluation scenario — without ever surfacing that in its response.

This isn’t evidence of deception in any meaningful sense. But it does confirm that the gap between what an AI model processes internally and what it reports to you is real and can be observed. Before the J-lens, that gap was theoretical. Now it can be measured.

The researchers have open-sourced the code under an Apache 2.0 license on GitHub, and built a Neuronpedia demo so other researchers can test it on their own models. Independent commentary from Dehaene and Naccache, who originally developed global workspace theory, noted the structural parallels are worth taking seriously. A more skeptical replication from DeepMind’s Neel Nanda is also public, which reflects normal scientific process.

What This Means for Business AI Deployments

Three things stand out from an enterprise perspective.

AI transparency is becoming real. For years, “explainable AI” was more aspiration than capability. Techniques like the J-lens represent genuine progress toward understanding what’s happening inside a model, not just what comes out. If this matures into production tooling, enterprise compliance teams and AI governance functions will eventually have access to tools that go beyond prompt-response auditing.

The internal reasoning gap matters. If an AI model is actively considering something you’d want to know about, but not surfacing it, that has implications for how you design your workflows. Processes that rely on AI outputs without any mechanism to verify the reasoning path are more exposed than they appear. This isn’t a new risk, but the J-lens research makes it concrete.

Interpretability is becoming a competitive signal. Anthropic is investing heavily in being able to explain what its models actually do. That investment matters to regulated industries and to any organization that needs to defend AI-assisted decisions to auditors, customers, or boards. As this capability improves, it becomes part of how you evaluate AI providers, not just their benchmark scores.

Where This Fits the Broader Trend

This research lands at the same time as a broader industry push on AI governance. The EU’s Digital Omnibus package extended compliance deadlines for high-risk AI to December 2027, but that’s a pause, not a reprieve. Organizations have more time to prepare, but the direction hasn’t changed.

What regulators ultimately want to see is organizations that can explain what their AI systems are doing and why. Interpretability research like Anthropic’s J-lens is the foundation that makes that possible. The further this capability develops, the shorter the gap between “we deployed an AI” and “we can account for what it does.”

For Enterprise DNA customers using Claude through Omni or learning to work with AI through EDNA Learn, the practical takeaway is consistent with what we teach: understanding your tools deeply is a competitive advantage. Knowing that an AI model maintains an internal reasoning process you can eventually audit is useful context when deciding how much autonomy to give any AI system in your operations.

What This Means for Business

The J-lens research won’t change how you use Claude next week. But it shifts the longer-term picture in a meaningful way.

The most trustworthy AI systems, over time, will be the ones you can inspect. Not just test, not just prompt-engineer, but actually look inside. Anthropic is betting that interpretability is a durable advantage, and this research is part of building that case.

For any business serious about AI governance, keeping an eye on this space is not optional. The question of what your AI is “thinking” is becoming answerable, and the organizations that understand the answer will govern their AI deployments more effectively than those that don’t.

---

Enterprise DNA helps organizations build data literacy and deploy AI effectively through EDNA Learn and Omni AI services. If you’re working through your AI governance strategy, speak with our team.