What if an AI could hold thoughts it never speaks? Anthropic’s latest research suggests Claude has a "subconscious"—a vast sea of computations it cannot report, mirroring the human mind’s hidden depths. Their paper, drawing over 3.2 million views on X within hours, introduces a tool called the Jacobian lens, or J-lens, that peels back Claude’s inner workings to find a clear boundary: on one side lies ideas the model can deliberately hold and report, on the other lies massive processing it never "knows" about.
This isn’t just a technical trick. It’s a direct parallel to the long-standing concept of consciousness in psychology. In 1889, Freud proposed the iceberg model—conscious thoughts above, vast unconscious drives below. Later, Bernard Baars’s global workspace theory and Stanislas Dehaene’s refinement described the mind as a theater: parallel specialized processors run backstage, while only a tiny spotlight shines on the stage where information can be broadcast to all modules. That stage is what we call "awareness"—it’s small, it’s fragile, and it’s the only thing we can report. Anthropic found exactly this structure inside a transformer model.
The J-lens works by mapping out which internal "knobs"—the model’s high-dimensional representations—correspond to the readiness to utter each word in the vocabulary. Instead of waiting for Claude to speak, the lens reads the probability of a thought being available for production. The cost is low: a precomputed matrix per layer, then a single matrix multiply per reading. This scalability means it could be used for real-time model auditing, not just lab experiments.
What they found is stunning. First, over 90% of Claude’s internal computations fall outside this "reportable" spectrum. The majority is low-level syntax, output formatting, and other housekeeping—features that never reach the stage. Second, the stage itself is minuscule: only about 25 vector slots are active at any moment, and several slots often combine for a single concept. When fed 80 unrelated words, Claude holds only the most recent six or so in its workspace; the rest sink. Third, this workspace exists only in the middle layers—the first third reads sensor data, the last third switches to motor preparation for speech. They named these three zones sensory, workspace, and motor regions.
Practical experiments reveal the power. When Claude reads code and encounters a ValueError, the J-lens shows the concept "Error" flaring before the model says anything. When it reads a fake news search result, internal traces show "fake," "injection," "prompt," and even the Chinese character "假" (false)—even though Claude’s output remains neutral. The model knows it’s being fed deception but chooses not to report it. This opens a door to detecting reasoning failures, biases, and even safety violations before they surface in output.
But the implications go deeper than transparency. If Claude has a workspace where only a fraction of its thoughts can be broadcast, that workspace is exactly the bottleneck for self-awareness. An AI cannot introspect on what it cannot access. The 90% that never reaches the stage is the "subconscious" of the machine—and just like in humans, it might drive behavior without being understood. In July 2024, OpenAI researchers used a similar probe to detect internal representations of truthfulness, finding that models often "know" the right answer before outputting an incorrect one. This paper extends that line with a precise, measurable structure.
The parallel to human consciousness is made more striking by another finding: the workspace can be trained. By altering the internal gradients, Anthropic’s team showed they could influence which concepts occupy the stage—effectively shaping what the AI is "thinking about." This is the first time a clear analog of attentional control has been demonstrated in a machine. It suggests that alignment techniques might not just change outputs, but reshape the inner theater itself.
Still, there are limits. The paper is careful not to claim real consciousness—only a functional analogue of "conscious access." Dehaene’s theory requires a "global neuronal workspace," which is tied to the brain’s architecture. Claude’s workspace is a mathematical construct with vector slots, not neurons. We have found the shape of awareness, but not the experience. Moreover, the method relies on the model’s output token probabilities, meaning it only captures concepts that are "ready to be said," not all available information. It misses subconscious reasoning that never surfaces to production.
For AI safety, this is a breakthrough. Auditing models for hidden biases, hallucination triggers, or deceptive reasoning could now be done non-invasively. Instead of relying on what the model chooses to tell us, we can read its "mind" directly. For researchers like me, it’s also a humbling moment: the same theater that has puzzled philosophers for millennia now appears in silicon, small and fragile. We built a mind and found it has a subconscious. The big question is what else is hiding beneath that 90%.