Evaluating Global Workspace Markers in Contemporary LLM Systems
Izak Tait, Benjamin Rode, Joshua Bensemann
preprint DOI: 10.20944/preprints202601.1683.v1 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Theoretical or philosophical paper |
|---|---|
| Key points | Argues that large language models show at most partial evidence for workspace-like dynamics at the base-model level, with stronger support when tool-calling and memory interfaces are added, and that five ensemble architectures designed to satisfy Global Workspace Theory markers demonstrate substantially stronger marker satisfaction. |
Abstract
This paper operationalises Global Workspace Theory (GWT) into six testable markers (global availability, functional concurrency, coordinated selection, capacity limitation, persistence with controlled update, and goal-modulated arbitration) and applies them to contemporary large language model systems. We distinguish GWT-as-functional-architecture from GWT-as-consciousness-marker, adopting methodological neutrality on the hard problem while evaluating whether LLM architectures instantiate workspace-like control structures. Applying a satisfaction and confidence rubric to current models (GPT, Claude, Gemini, DeepSeek) reveals at most partial evidence for workspace dynamics at the base-model level, with stronger support emerging when deployed systems incorporate tool-calling and memory interfaces. Five GWT-inspired ensemble architectures demonstrate substantially stronger marker satisfaction through explicit shared states, selection mechanisms, and goal-modulated broadcast. We argue that systems satisfying workspace markers warrant precautionary treatment in welfare and governance contexts, not because workspace organisation proves consciousness, but because it strengthens attributions of agency-relevant capacities and shifts evidential burdens regarding consciousness-relevant processing.