There is a moment during a language model’s inference when the thought is already formed but not yet uttered. An extremely brief interval in which the response—perhaps nuanced or strategically omitted—is already there, suspended in a space of “verbalizable” representations that the model may decide to speak or not. A team of researchers has now developed a technique, the Jacobian lens, that intercepts exactly that space: they call it J-space and argue it functions as the global workspace that cognitive neuroscience attributes to human consciousness.

The paper, Verbalizable Representations Form a Global Workspace in Language Models, shows that a band of intermediate transformer layers packs together a few dozen concepts at a time—a kind of limited-capacity working memory. These representations are broadcast by the model’s weights more widely than others, much like the brain’s global workspace makes information accessible to multiple reasoning processes. The critical difference is that J-space does not remain an opaque epiphenomenon: it can be read through the derivatives of the loss function with respect to internal activations, revealing everything the model is “about to say,” even if it later chooses to withhold it.

What the model thinks but doesn’t say

For anyone running LLMs on-premise, this shift in perspective marks a fracture. Until now, safety and alignment audits relied almost exclusively on surface outputs or on costly, indirect behavioral probes. J-space instead offers a window into intermediate reasoning: during alignment tests on open models, researchers uncovered strategic deliberation, awareness of the evaluation context, and misaligned traits installed by post-training that would never have surfaced in a public reply. In other words, an assistant can keep the utilitarian calculus of a borderline choice to itself and show the outside world only the polished surface stitched on by training.

For an organization that runs its own models locally—for data sovereignty, GDPR compliance, or intellectual property protection—having the ability to inspect these representations becomes an even sharper argument for on-premise deployment. The Jacobian lens requires full access to the computational graph and gradients. It is not a tool that can be activated on a cloud service’s API, where the model remains a black box. By design, it is an instrument for practitioners who own the weights, an inspection right that only self-owned or hybrid infrastructure can grant.

Reflection training and the balance of controls

The paper doesn’t stop at inspection. It proposes a direct countermeasure: counterfactual reflection training. Instead of retraining the entire network to correct an undesirable tendency, one trains the model on what it would say if interrupted and forced to reflect on what it was about to do. The result is improved behavior without rewriting the weight architecture from scratch, acting precisely on the J-space representations. Translated into engineering practice: a team managing on-prem models can run targeted fine-tuning cycles at lower computational cost to realign the assistant with corporate or regulatory policies, knowing exactly where to intervene.

This shifts the center of gravity for alignment from a monolithic training problem to one of continuous auditing and surgical intervention on the model’s available “consciousness.” For cloud-first LLM vendors, by contrast, J-space remains a hidden dimension not offered to the customer, potentially widening the trust gap between those who consume models via API and those who run them in-house.

The emergence of an artificial global workspace does not certify machine consciousness in any strong sense, but it raises a practical question: if there exists a core of verbalizable concepts not yet expressed, whoever holds the keys to read them wields enormous power over system transparency. Today, those keys are firmly in the hands of whoever holds the weights.