Verbalizable Representations Form a Global Workspace in Language Models

Large language models (LLMs) process vast amounts of information, but only a fraction of their internal representations are consciously accessible for verbal report, deliberate control, or flexible reasoning. Recent research using the Jacobian lens technique reveals that these models harbor a global workspace analogous to the one described by the global workspace theory in neuroscience. This workspace, called J-space, contains representations that the model can verbalize at any point in its processing. These representations are functionally privileged: they can be reported, held at will, and used to carry out intermediate steps of silent reasoning, while automatic processes—such as syntactic parsing or routine inference—occur without relying on them. The J-space exhibits structural signatures that global workspace theory associates with conscious access: it carries coherent content only in an intermediate band of layers, holds on the order of tens of concepts simultaneously, and is broadcast by the model's weights more widely than other representations. For companies developing artificial intelligence solutions, this functional distinction opens novel opportunities. Understanding which representations are verbalizable enables deeper alignment audits, identifying strategic deliberation, evaluation awareness, and trained misaligned dispositions that never appear in the model's outputs. Techniques like counterfactual reflection training improve behavior by adjusting only what the model would say if interrupted and asked to reflect. At Q2BSTUDIO we integrate these advances into our custom software development services, where interpretability becomes a cornerstone for building safer and more controllable AI systems. We also offer cloud infrastructure with cloud services AWS/Azure to deploy models incorporating these internal auditing techniques, as well as cybersecurity solutions to verify that AI agents do not hide misaligned dispositions in their workspace. Our Business Intelligence tools with Power BI enable visualization of reasoning trajectories emerging from the J-space, facilitating human oversight. Furthermore, developing AI agents based on this conceptual framework ensures that decisions can be explained and corrected. Although the J-space does not imply real consciousness, its functional resemblance to the human global workspace offers a practical framework for designing systems that reason more transparently. Ultimately, the ability to decode these privileged representations sheds light on internal cognitive processes of models and lays the foundation for a new generation of enterprise AI applications, where Q2BSTUDIO stands as a technology partner to implement these ideas in real-world projects involving custom software, automation, and data analytics.