On July 6, 2026, Anthropic’s interpretability team published “A global workspace in language models,” describing an emergent internal subspace inside Claude that the authors call the J-space. Using a new method they name the Jacobian lens, or J-lens, the researchers map, for each token in the vocabulary, the internal activity pattern that makes the model more likely to say that word later rather than immediately. The result is a privileged set of representations that behaves like the “global workspace” long theorized in cognitive science as the basis for conscious access.
The paper reports that the J-space holds only a few dozen concepts at a time and accounts for less than ten percent of overall neural activity, yet plays an outsized role in higher-order cognition. Claude can accurately report what is in its J-space when asked what it is thinking about, can deliberately modulate the contents when told to hold a topic silently, and reuses single concepts flexibly across downstream tasks. When the researchers ablate the J-space, fluent language generation and basic factual recall survive, but multi-step reasoning, summarization, and other compositional abilities collapse.
Anthropic open-sourced an implementation of the J-lens and partnered with Neuronpedia for an interactive demonstration on open-weight models. The team also points to safety applications, including detecting when a model recognizes it is being evaluated, catching fabricated data, and surfacing hidden goals in deliberately misaligned test models. The work drew wide commentary, including a more skeptical replication effort from outside researchers.
For business and technical leaders, the finding matters because it shifts interpretability from studying model outputs to reading an internal, causally important layer of “thought.” That is directly relevant to auditing, monitoring, and trusting AI systems in production.