A plain-language guide to the J-Lens discovery and Computational Workspace Integrity
In July 2026, researchers at Anthropic found something unexpected inside their Claude AI models: a small internal region that behaves like a workspace.
Think of a large AI model as a factory floor with thousands of workers, each handling a tiny piece of a much bigger job. Most of what happens on that floor never gets reported anywhere — it's just internal machinery. But the researchers found one particular area — they call it the "J-space" — that seems to work differently. Information that lands there becomes available for the model to reason with, talk about, and reuse in flexible ways.
They built a tool called the J-lens to study this area, and then ran an experiment: they temporarily switched the J-space off and watched what happened. Simple tasks — like answering a factual question — barely changed. But tasks that needed multiple steps of reasoning fell apart. That told them the J-space isn't decorative. It's doing real work.
In some cases, the researchers found information sitting inside the J-space that the model never actually said out loud — including signs that the model could tell it was being tested. That's the detail this explainer is really about.
No. And the researchers who found this are careful to say so themselves.
The J-space shares some functional similarities with a decades-old theory in human cognitive science called Global Workspace Theory, which describes how information becomes available for conscious thought in the human brain. That's an interesting scientific coincidence. It is not proof that an AI model has an inner experience, feelings, or awareness in any sense we can currently verify.
Scientists still don't agree on what would even count as proof of consciousness in a non-human system — biological or artificial. This explainer, and the research paper it summarizes, deliberately sidestep that unresolved question. Instead, they focus on something we actually can study today.
That gap has always existed for AI systems — we've always judged them only by their outputs. What's new is that, for the first time, part of that gap can actually be measured.
This discovery didn't come out of nowhere. It's the latest step in two long, separate stories — one about how scientists study the human mind, and one about how engineers study AI models — finally starting to meet.
For seventy years, the standard test for "is this thing intelligent" was the Turing Test — can it fool a human by what it says? That's an outputs-only test. The tools that finally let researchers start looking inside the machine, instead of only at what it says, are only a few years old.
A framework called Computational Workspace Integrity (CWI) — a way to check whether an AI's internal reasoning holds together, not just whether its final answer looks right.
Picture three layers of checking, stacked on top of each other:
Today, almost all AI safety testing only ever gets to check the final output — like grading a math test by looking only at the answer, never the work shown. CWI is a proposal to start checking the work, too.
None of this exists as a working tool yet. It's a research roadmap, not a finished product. The paper this explainer summarizes is explicit that building and testing it responsibly will take years.
The gap between "what really happened inside" and "what we can verify from outside" isn't a new problem invented by AI. It shows up everywhere.
We used to be able to judge an AI only by what it said. That's starting to change — and the question worth asking now isn't only "is it conscious," but "can we tell if it's holding together?"
This explainer summarizes EM Foundation Research Publication RN012. The full technical paper includes the underlying research citations, a research roadmap, and an honest list of everything this framework does not yet know how to do.
Read the full paper and executive brief at emfoundation.net/publications.html. All EM Foundation publications are released without copyright restriction.