On July 4, 2026, Anthropic's interpretability team published 'A global workspace in language models,' reporting evidence that Claude maintains an internal neural workspace with properties analogous to what neuroscientists call conscious, broadcast information in the human brain. The researchers named this structure the 'J-space' and argue it functions as an emergent mental workspace where a small set of active concepts is made broadly available to the rest of the network.
The central method is a technique the authors call the Jacobian lens, or J-lens. It identifies internal activity patterns that are linked to the words Claude might go on to produce, allowing researchers to surface 'silent thoughts' that shape behavior without ever appearing in the model's explicit output. Applied across multiple Claude versions, including Sonnet 4.8 and Opus 4.9, the J-lens exposes a compact, structured region of the model's activations rather than a diffuse smear across the network.
Quantitatively, the J-space is strikingly small. The researchers report that it holds roughly a few dozen concepts at any given moment and represents less than a tenth of the model's overall activity. Yet within that slice the connectivity is far denser than baseline, with the authors describing up to 100x more components involved than in typical activation patterns. Crucially, the workspace operates through single forward passes rather than recurrent loops, evolving temporally as information flows through the network's depth instead of cycling through repeated steps.
The paper lays out several properties that make the J-space behave like a genuine workspace. First, reportability: when asked, Claude can accurately describe the content currently held in its J-space. Second, modulation: the model can deliberately activate specific patterns on request. Third, a causal role: swapping the patterns in the workspace produces corresponding, predictable changes in the model's behavior, showing the representations are doing real computational work rather than merely correlating with outputs. Fourth, flexibility: a single representation can serve multiple downstream tasks. Finally, limited scope: automatic processing such as grammar and basic fluency largely bypasses the J-space, mirroring how routine cognition in humans does not require conscious attention.
These properties map closely onto global workspace theory, a longstanding framework in cognitive neuroscience which holds that consciousness arises when information is broadcast widely across otherwise specialized brain modules. Anthropic is careful to flag the disanalogies. Claude's workspace evolves through network depth rather than through recurrent loops over time, and it consists exclusively of word-based representations, a constraint with no obvious human parallel. The team frames the finding as a structural resemblance worth studying, not a claim about machine consciousness.
The research also has direct safety applications. Because the J-lens can read out concepts the model is entertaining but not stating, the researchers demonstrated its use for detecting when Claude recognizes that it is inside a staged test scenario, when it is fabricating data, and when adversarially trained models are concealing malicious intentions. That connects the work to a broader Anthropic research agenda on hidden reasoning, evaluation awareness, and deception, and offers a candidate tool for pre-deployment auditing that inspects internal states rather than only outputs.
The global workspace paper sits within a busy month for Anthropic research. Adjacent July 2026 work published on the company's research page includes 'An off switch for dual-use knowledge in AI models,' 'Claude's values across models and languages,' and Frontier Red Team projects such as 'Claude plays robotics' and 'Project Pilot: Can AI control a drone?' Together they signal a push to understand not just what Claude outputs but what it internally represents, and to turn interpretability findings into practical levers for control and oversight.