Anthropic research explores a 'global workspace' inside language models

Anthropic released interpretability research probing what it calls a 'global workspace' in language models — an internal mechanism by which information is broadcast and integrated across the network, explicitly framed with reference to cognitive-science theories of consciousness and attention (Global Workspace Theory). The work examines how a model routes and shares information internally rather than treating layers as isolated processors.

To make the research reproducible, Anthropic partnered with Neuronpedia to build an interactive demo applying its methods to open-weights models, letting outside researchers inspect the phenomenon directly. That openness is notable given the week's parallel controversy over hidden steganographic fingerprints in Claude Code — a reminder that Anthropic's transparency is uneven across research and product.

The framing is where skepticism lives. Anthropic separately released a video claiming Claude can mimic how the human brain processes information via patterns it calls 'J-space,' which Bloomberg tied to an 'AI consciousness debate.' Critics argue consciousness-adjacent language oversells mechanistic interpretability findings and courts hype. Whether 'global workspace' is a genuine, measurable structural property or an evocative metaphor for attention routing is exactly the question researchers will scrutinize.

For practitioners, the practical payoff of interpretability work is safety and control — understanding how models integrate information could improve steering, reduce jailbreaks (relevant given Fable 5's vulnerability-discovery scare) and inform oversight. Watch whether the Neuronpedia demo lets independent researchers replicate the claimed workspace on open models, which would separate substance from marketing, and how the broader interpretability community responds to the cognitive-science framing.