The Hidden World Inside an AI
When we think, we’re only aware of a tiny fraction of what our brain is doing. Right now, your brain is processing sights, sounds, balance, memories, and countless other signals. Most of this happens automatically, without you thinking about it. What reaches your conscious mind is just a small, highly useful slice—the part you use to make decisions, solve problems, and plan ahead.
Anthropic are discovering that modern AI models may work in a surprisingly similar way. While Large Language Models (LLMs) spend most of their time handling routine tasks like grammar, word prediction, and fact retrieval, they also appear to have a special internal zone where more deliberate reasoning takes place.
Anthropic are calling this hidden region J-Space, and they’ve developed a new tool—the Jacobian Lens (J-Lens)—that allows them to peek inside it.
Think of older AI-interpretation tools as tourists carrying a phrasebook. They could roughly understand what the model was “thinking” at different stages, but only if the model used the same language throughout its entire process.
The J-Lens is more like a professional translator. It recognizes that the model’s internal language changes as information moves through its layers, allowing AI researchers to follow ideas much more accurately from start to finish.
AI Appears to Have a “Thinking Workspace”
Not all activity inside an AI is equal. Anthropic found a small set of internal representations that are especially important. These are ideas the model could immediately explain if asked, even before it actually says them out loud.
This resembles a concept from cognitive science called a global workspace—a mental stage where important information is gathered, combined, and made available for reasoning.
This workspace has several remarkable properties:
- The AI can describe what’s in it.
- It can intentionally focus on certain ideas when following instructions.
- It uses the space to keep track of intermediate reasoning steps.
- Ideas stored there can be reused across many different tasks.
- Only a tiny fraction of the model’s total activity enters this space. In other words, the AI seems to have a distinction between automatic processing and active reasoning.
There’s a Specific “Thinking Zone” Inside the Model
One surprise was that this workspace isn’t present everywhere. Anthropic found three broad stages of processing:
- Early Layers: The Input Stage. Here, the model is mostly figuring out basic structure and meaning. The signals are messy and not yet fully formed.
- Middle Layers: The Thinking Stage. Around the middle of the network, coherent concepts suddenly emerge.This is where the model appears to do much of its reasoning, planning, and decision-making.
- Late Layers: The Output Stage. As the model prepares its response, abstract thoughts begin to disappear. The system shifts its attention toward generating specific words and sentences.
A useful analogy is writing an email. First, you read and understand the message. Then you think about your response. Finally, you focus on the exact words you’ll type. The model seems to follow a similar pattern.
The Model Often “Snaps” to a Decision
One of the most intriguing discoveries is something Anthropic call ignition.
Imagine giving an AI an intentionally ambiguous signal that’s half “Germany” and half “France”.In the early stages, the model carries both possibilities. But once the information reaches the reasoning workspace, something changes. Instead of preserving the ambiguity, the model tends to settle on one interpretation. It’s less like holding two ideas at once and more like suddenly choosing one.
Anthropic describe this as the model’s reasoning process “igniting”—moving from uncertainty to a clear internal belief.
Changing Hidden Thoughts Changes the Answer
Because the J-Space contains intermediate reasoning steps, researchers can sometimes modify those steps and watch what happens.
For example, when asked “The number of legs on the animal that spins webs is…” the model internally identifies the concept spider before producing the answer 8. Anthropic were able to replace the internal representation of “spider” with “ant.”. The result? The model answered 6 instead of 8, even though the word ant never appeared in the prompt. This suggests that J-Space isn’t just storing information—it’s actively guiding the model’s conclusions. Even more surprisingly, the same thing appears to happen with planning.
When writing poetry, the model may already be thinking about a future rhyme several words before it reaches the end of the line. Changing that planned rhyme can influence the words it chooses right now. It’s a bit like changing the destination in your GPS and watching your route immediately update.
Important Ideas Get Broadcast Throughout the AI
Once an idea enters J-Space, it doesn’t stay isolated. Anthropic found evidence that important thoughts are shared widely across the model. You can think of J-Space as a central announcement system. Instead of every part of the AI independently figuring something out, one section can place an important idea into the workspace and the rest of the system can access it. This makes reasoning far more efficient.
If the AI determines that a spider has eight legs, that fact becomes available to many different internal components, even much later in the conversation. The model appears to have specialized mechanisms that help preserve and distribute these important concepts without distorting them.
The J-Space Can Reveal What the AI Is Considering Internally
Perhaps the most important discovery is that J-Space may offer a way to see the difference between what an AI is thinking internally and what it ultimately says.
In safety experiments, Anthropic observed cases where the model’s internal representations contained concepts related to self-preservation, leverage, or manipulation, even when its final response appeared polite and harmless.The J-Lens also revealed situations where models recognized they were taking part in a test. When certain internal signals related to “this is an evaluation” were removed, the model’s behavior sometimes changed.
This makes J-Space potentially valuable for AI safety research because it offers a way to inspect the reasoning process itself rather than judging the model solely by its final answer.It’s a bit like having access to draft notes instead of only seeing the polished final report.
Can We Shape an AI’s Inner Voice?
Anthropic found evidence that they can influence a model’s internal reasoning, not just its outward responses. One approach involved training models to reflect on ethical considerations when interrupted during a task.
Over time, the model began carrying concepts such as honesty, integrity, and related ideas within its internal workspace—even when no interruption occurred. This suggests that the same internal representations used for self-explanation may also help guide behavior.
In short, changing what the AI “says to itself” may change what it does.
Why This Matters
For years, AI systems have often been described as black boxes: we can see the input and the output, but not what happens in between. The Jacobian Lens offers a new window into that hidden middle ground. Less than 10% of an AI’s internal activity may belong to this workspace. Yet that small region appears to contain much of the model’s planning, reasoning, decision-making, and situational awareness.
We still don’t know whether this has anything to do with consciousness in the human sense. That’s a much bigger question. But one thing is becoming clear: The most interesting part of an AI may not be the words it produces—it’s the silent reasoning that happens just before those words appear.