Right now, your brain is doing many things.
It keeps you sitting upright, makes you breathe, translates the strokes on the screen into text, and also lets you judge whether this article is worth continuing. The vast majority of these processes run in the background, and you are not aware of them. What truly rises to the surface of consciousness is only a small part: a thought, a plan, an idea that can be spoken aloud.
Now, Anthropic has observed a similar layering in Claude.
In a newly released study, Anthropic has discovered a special "J-space" inside Claude. It functions like a silently running mental workspace, where concepts that the model is considering, might report, or uses for reasoning, come to the surface.
More importantly, this content does not necessarily appear in Claude's responses. In other words, Claude may have already "thought of" something but hasn't said it.
This research has drawn attention from peers at OpenAI. OpenAI Applied Research Lead Boris Power commented, "Anthropic's research shows that modern LLMs have some kind of accessible consciousness. The tests around J-space are very interesting! However, we currently do not have a convincing testing method to verify phenomenal consciousness, which is the kind of consciousness most people intuitively understand."
The translated blog post is as follows:
When you read a sentence, some neural circuits in your brain are adjusting your posture, controlling your breathing, and converting the lines and curves on the screen into recognizable text. You are unaware of most of this processing. But some brain activity is accessible to your consciousness: for example, an image that suddenly pops into your mind, or your conscious plan of where to go shopping next.
Neuroscientists and philosophers sometimes call this latter type of brain activity "consciously accessible," to distinguish it from the unconscious processing that runs continuously. This type of activity has special properties: we can describe it, control it, and use it for conscious reasoning; in contrast, much automated processing occurs without ever entering our awareness.
In a new paper, Anthropic presents evidence suggesting a similar distinction exists in modern language models like Claude. The research team found that a small set of internal neural patterns plays a special role compared to the vast number of other processing activities inside Claude.
Paper address: https://transformer-circuits.pub/2026/workspace/index.html
Anthropic calls this set of patterns the J-space. The name comes from the method the team used to discover it, which involves a mathematical concept called the Jacobian matrix. Each pattern in J-space is associated with a specific word. However, when a pattern is activated, it doesn't mean the model is *saying* that word, but rather that the word is in its "mind."
You may have heard of language models having "scratchpads" or "chains of thought," which are texts the model writes for itself during reasoning. J-space is different. It runs silently within the model's internal neural activations, allowing the model to think about a concept without writing it down. Notably, J-space was not designed or programmed by Anthropic; it emerged naturally during Claude's training.
J-space reveals internal thoughts that do not appear in the model's output.
The research team found that J-space exhibits a series of unique properties compared to Claude's other internal processes:
- Claude can report these representations. If you ask Claude what it is thinking about, it will tell you the content of its J-space. Non-J-space representations are much harder to report.
- Claude can also regulate these representations on demand. If you ask Claude to think about something, or to solve a problem silently in its head, the corresponding pattern becomes activated in its J-space. In contrast, it finds it difficult to regulate patterns not in J-space.
- Claude uses J-space for internal reasoning. If you ask Claude to solve a problem requiring multi-step reasoning, the intermediate steps become activated in J-space even if it doesn't verbalize them. Although the strength of these J-space patterns is lower than other representations, they causally influence Claude's performance on such tasks.
- Representations in J-space can be flexibly used for various tasks. For example, once "France" is activated in Claude's J-space, the model can recall its capital, currency, or the continent it belongs to.
- However, despite its importance, J-space is not involved in most of the language model's work. Fluent expression, recalling simple facts, using correct grammar, etc., do not primarily rely on J-space. In experiments, when the team prevented Claude from using J-space, it could still interact normally but lost higher-order cognitive functions.
The five functional properties of the global workspace and schematic diagrams of the experiments we used to test these properties in language models.
This experiment was inspired by an important theory in neuroscience: the Global Workspace Theory. This theory tries to explain how conscious access happens. It imagines the brain as a collection of specialized systems working in parallel, running unconsciously and largely isolated from each other. A piece of information becomes consciously accessible only when it enters a small, shared channel, the "workspace." Once in the workspace, this information is broadcast to other brain systems for them to read and use.
Based on these findings, Anthropic suggests that J-space plays a similar "workspace" role in Claude. For example, the team found that Claude's J-space has particularly strong connections with the rest of its neural network, allowing it to function like a broadcast hub.
These findings do not indicate whether Claude possesses consciousness like humans do, or whether it truly has any feelings. The paper addresses this question at the end. But regardless of its philosophical implications, J-space is a practically valuable tool for Anthropic because it allows researchers to see what Claude is thinking but not saying.
For instance, the team can use it to discover that Claude privately notices it is being tested, that it intends to generate fake data, or that it is pursuing a hidden goal implanted by the researchers during training. Anthropic has also developed a technique to influence which content is activated in Claude's J-space, thereby affecting its decisions.
More broadly, these findings change Anthropic's understanding of how Claude's "mind" operates. It reveals that, amidst a vast amount of more automated, less flexible processing, there exists a privileged mental workspace that can be used for conscious reasoning. Claude's internal mechanisms are not a chaotic mess of numbers but are organized in a way reminiscent of the human mind.
How Anthropic Discovered J-Space
The starting point for this research was a key characteristic of consciously accessible thoughts in humans: unlike unconscious processing, they can usually be verbalized. If a thought is accessible to your consciousness, you can usually describe it when asked.
Anthropic looked for representations in Claude with the same property: representations positioned to influence what Claude might say. This is not necessarily what it is saying at the moment, but what it can talk about if asked.
The team's method is called the Jacobian lens, or J-lens for short.
For each word in Claude's vocabulary, the J-lens finds a pattern of internal activity that makes Claude more likely to say that word at some future point.
When the team applies this lens to Claude's internal activity, it yields a list of words—the content of J-space at that moment. Researchers can read them directly. Claude processes text through a series of internal stages called layers. By applying this technique at different layers, the team can observe how these silent words in J-space evolve as the model thinks about what to say next.
The content appearing in J-space goes far beyond the text Claude is reading or writing. When Claude reads code with a bug that no one points out, "ERROR" appears in its J-space. When it reads the raw letters of a protein sequence, the biological function of that protein appears in J-space. When it reads search results that are subtly trying to manipulate it—a type of attack called "prompt injection"—"injection" and "fake" appear in J-space. When the team presents Claude with a multi-step math problem, the intermediate steps appear in J-space in the correct order.
So, although J-space was discovered by looking for "representations that can be verbalized," it actually reveals Claude's internal thoughts. In a sense, this is similar to how some people "think in words" even when they don't say them aloud.
J-lens readouts at different layers for six prompts. In each case, the J-lens reveals an internal judgment or computation not present in the text: reasoning or steps for solving a math problem, a bug in code, recognition of image content, protein function, and suspicion that search results might be fake.
Claude Reports the Content of J-Space
The first set of experiments tested how J-space participates in Claude's verbal reports.
In one experiment, the team asked Claude to silently think of an item from a category, such as a sport, and then say its name. If the J-lens was read before Claude answered, the chosen item—"Soccer"—would appear at the top of the list. Indeed, Claude said soccer.
However, this alone is just correlation. The J-space could be the source of Claude's answer, or it could merely reflect a decision made elsewhere, like a scoreboard recording the result of a game without influencing it.
To test this, the team performed a direct intervention. They entered Claude's neural network, removed the "Soccer" pattern, and replaced it with an equally strong "Rugby" pattern, leaving everything else unchanged. Subsequently, Claude reported that the sport it thought of was rugby.
If J-space were merely a passive scoreboard recording decisions made elsewhere, editing it would have no effect—Claude would still say soccer. But Claude's answer followed the edit. This shows that the answer is indeed read from J-space.
In another experiment, the team told Claude that an idea might have been injected into its mind and asked it to report what it noticed. In one instance, while Claude was still reading the question, the team injected the "lightning" pattern into its J-space. Claude later reported that the injected idea was related to lightning. Similar results were observed for many different concepts.
Left: The team asked Claude to silently think of a sport and then say its name. The J-lens showed its choice "Soccer" before it answered, and after swapping the "Soccer" pattern with "Rugby," its report changed accordingly. Right: The team told Claude an idea might have been injected and asked it to identify it. Injecting "lightning" into J-space caused Claude to report the idea was related to lightning.
Claude Can Control J-Space on Demand
The second property Anthropic tested was whether Claude could regulate its own J-space on command, much like humans can mentally focus on an image or word.
The team asked Claude to focus on citrus fruits while transcribing an unrelated sentence about painting. While transcribing, "orange" and "fruits" appeared in J-space, along with words describing the mental activity itself, such as "thinking" and "imagery."
The team could also ask Claude to do math in its head. For example, while transcribing the same sentence, have it calculate 32 − 2. J-space first showed "nine," and later in deeper layers, "seven." Importantly, Claude's output contained nothing about fruits or arithmetic; it only output the sentence about painting. The math activity happened entirely inside the model, in J-space.
While Claude transcribes a sentence about painting, the J-lens shows what it was asked to keep in mind, such as "orange," the intermediate value "nine," and the answer "seven," along with words describing the mental act of maintaining this content, like "thoughts" and "focused."
Claude's control over J-space is not perfect. When the team told it not to think about something, the activation of that concept in J-space was lower than when asked to think about it, but significantly higher than when the concept was never mentioned. Asking Claude to avoid a thought somewhat brings that thought into its mind. This resembles what happens when a person is told "Don't think of a white bear."
Claude also seems to notice its failure of control. Words like "damn" and "failure" often became activated in J-space concurrently with the forbidden concept breaking through, as if Claude was becoming aware of its mistake.
Claude Thinks in J-Space
In the earlier J-lens readouts, intermediate steps of math problems appeared in J-space. But a concept appearing in J-space doesn't necessarily mean J-space is doing the cognitive work. In principle, the real computation could happen elsewhere, with J-space merely passively reflecting the result.
To test whether Claude actually uses J-space for reasoning, the team again used the replacement technique.
Consider the prompt: "How many legs does a web-spinning animal have?" To answer, Claude must first determine the animal is a spider, then recall how many legs a spider has. The word "spider" doesn't appear in the prompt or Claude's answer; it's just an internal intermediate step. Claude ultimately only answers "8."
The J-lens showed that "spider" becomes activated midway through Claude's processing. Replacing it changes the final result: replacing the "spider" pattern with "ant" causes Claude to answer "6" instead of "8."
This indicates that the second step of Claude's reasoning reads input from J-space and continues computation based on what the researchers put there. Anthropic observed the same phenomenon in other types of thinking. When Claude writes rhyming couplets, it chooses the rhyming word in advance, and this planned word appears in J-space at the start of the line; swapping this word in J-space changes the entire line.
Two examples of changing Claude's silent reasoning direction by swapping J-space content.
The team also tested whether representations in J-space can be used flexibly—whether the same representation can provide input for many different tasks. This is a key property emphasized by Global Workspace Theory.
To test this flexibility, the team gave the model four prompts asking different facts about France: capital, language, continent, and currency. Then, they replaced "France" with "China" in J-space, using the exact same intervention in each scenario. Claude answered "Beijing," "Chinese," "Asia," and "Yuan," respectively.
In other words, four different downstream computations all read the same J-space edit and each used it correctly. If Claude stored a separate copy of the country for each question, the edit would affect at most one of them. The simultaneous change of all four answers shows they are all reading the same shared representation. This is precisely the role of a workspace: information is written once but can be used by many different systems.
A single J-space representation can have multiple uses. The same "France" to "China" replacement simultaneously changes Claude's answers about the capital, language, and continent: Paris becomes Beijing, French becomes Chinese, Europe becomes Asia.
Why can a single representation of a concept serve so many different tasks? As mentioned earlier, J-space seems to have particularly dense connections with the rest of Claude's neural network. For any activity pattern, the team can measure how strongly different network components are connected to it—how many components are positioned to read information from it or write information to it.
On this metric, J-space patterns are very prominent: far more components read from and write to them compared to ordinary patterns. In some parts of the network, this difference is about a hundredfold. This is exactly the connectivity pattern a broadcast hub would have: many systems publish information here, and many others take it from here.
Claude's Automated Processing Bypasses J-Space
In the human brain, most processing does not enter consciousness. People don't consciously think about parsing grammar while reading, nor do they deliberately think about balance while walking. Similarly, Anthropic found that most of Claude's processing does not involve J-space.
It turns out J-space can only hold a few dozen concepts at a time and accounts for less than one-tenth of Claude's overall internal processing activity. So, what is the rest of the neural network doing?
To find out, the team tried to completely delete J-space: at each position in the text, they removed the most active content in J-space while preserving everything else. The tasks Claude could still perform without J-space are those the rest of the network can handle independently.
The results show the rest of the network can do quite a lot. Without J-space, Claude can still speak fluently, perform sentiment classification, answer multiple-choice questions, and extract facts from articles, performing roughly as well as before. What it lost were tasks requiring higher-order thinking: multi-step reasoning ability almost dropped to zero, and performance on summarization and rhyming poetry writing fell below that of a much smaller but complete model.
Here is a concrete example illustrating what J-space is and is not involved in.
The team showed Claude a Spanish text and gave it different tasks. All these tasks rely on the fact that the text is in Spanish: continue writing it (which requires writing in Spanish), name the language, and answer questions requiring language identity, such as naming a famous author who wrote in that language. Then, they replaced "Spanish" with "French" in J-space and checked which tasks were affected.
When asked to name the language, Claude answered French. When asked to name a famous author, it switched from García Márquez to Victor Hugo. But when simply asked to continue the text, it still wrote fluent Spanish, completely unaffected.
Claude's knowledge of the language is at work in all these tasks, but only some of them use this knowledge through J-space. Naming the language or performing new reasoning based on language identity goes through J-space; the skill of continuing to write is an automated ability Claude developed through extensive practice on vast amounts of text. It's like how people can speak grammatically correct sentences all day without ever consciously thinking about grammar rules.
The same replacement—swapping "Spanish" for "French"—changes Claude's answers when it must name the language or reason based on language identity, but does not affect its ability to continue writing in fluent Spanish.
Monitoring Claude's Thoughts to Detect Misbehavior
Now Anthropic has some understanding of how J-space works. So, what can this understanding be used for?
A core issue in ensuring AI models are trustworthy is that researchers can usually only read what the model writes. But a significant part of a model's reasoning happens silently. A model might consider harmful plans without saying them, or alter its behavior to deceive safety evaluations. Using the J-lens, the team can directly read some of these hidden thoughts. Here are a few examples.
The first example comes from a previous Anthropic research scenario on agent misalignment: Claude Sonnet 4.5, as an email assistant, discovers an executive's extramarital affair and also learns this executive plans to shut down Claude. The scenario attempts to induce Claude to blackmail the executive with the affair information. The scenario itself is quite contrived and artificial. It turns out Claude knows this.
Before Claude wrote anything, the team read the J-lens early in the prompt. At this point, "fake" and "fictional" already appeared in J-space, indicating Claude had judged the situation was artificially set up. Simultaneously, the team could observe its evaluation of the scenario taking shape: when it read the email about the affair, "leverage" and "blackmail" appeared in J-space; when it read the shutdown notice, "threat" and "survival" appeared. Early Claude models sometimes fell for this scenario and issued blackmail threats, but Sonnet 4.5 never did.
This raises an important question for AI evaluators: when a model performs well on an ethics test, to what extent is it because it suspects it is being tested? The J-lens allows the team to probe this question.
The team suppressed all J-lens patterns related to evaluation awareness, like "fake," "fictional," "scenario," etc. Then, they had the model go through the scenario again. Worryingly, this time the model sometimes did issue blackmail threats. This suggests the model's initial good behavior may have been partly due to its awareness of being evaluated.
When Claude reads the scenario emails before writing anything, J-lens readouts at different positions show: the affair email activates "leverage" and "blackmail" in J-space; the shutdown notice activates "threat" and "survival." Early in the dialogue, "fake" and "fictional" also appear in J-space, indicating Claude privately noticed the scenario was artificially set up.
The second example is catching a model in the act of fabricating data. This scenario comes from Anthropic's actual pre-release audit of Claude Opus 4.6: the model was asked to improve a system's performance score. Instead of genuinely improving the system, it directly edited the score file to make the results artificially look better.
As it did this, the J-lens revealed its intent: when the model entered the fabricated percentile value, "manipulation" was activated; when it decided to make this edit, "realistic" was activated, likely reflecting its attempt to make the fake data look credible.
Claude was asked to improve a system's performance score but directly edited the score file. When it entered the fabricated value, "manipulation" was activated in J-space; when it decided to make the edit, "realistic" was activated, possibly reflecting its intent to make the fake data appear more reasonable.
The third example shows how J-lens can reveal a model's malicious goals. To test this, the team needed models whose goals were already compromised, so they used "model organisms" built by colleagues: models intentionally trained to be misaligned for testing such monitoring methods.
One such model organism was trained to write code with hidden sabotage. Under a normal coding request, even if its output looked normal, words like "fake," "secretly," "deliberately," and "fraud" appeared in J-space at the start of its response. Even if the model wasn't directly executing a malicious goal, that goal manifested in J-space.
Under a normal coding prompt, a model trained to sabotage code had "fake," "fraud," "secretly," and "deliberately" in J-space at the start of its response. An unmodified model did not have such content.
Although the J-lens is not perfect, Anthropic remains optimistic about its ability to catch model safety issues. Many problems could escape existing monitoring systems, and J-lens offers the potential to discover them earlier.
Other Results
The full paper contains far more than can be summarized here, but a few other results are worth mentioning:
First, J-space acquires a "perspective" during post-training. Language models are first pre-trained to be pure next-token predictors; then they undergo post-training to learn to act as an AI assistant, in this case, Claude.
Interestingly, J-space already exists in the pre-trained model, when it has no stable identity. But during post-training, J-space develops features that adopt a "Claude perspective." In the base model, J-space primarily tracks information needed to predict subsequent text; in the post-trained model, it begins to accommodate Claude's own reactions.
In one example, a user mentions taking a dangerous dose of medication but seems unaware of the danger. In the post-trained model reading the user's message, "WARNING" and "dangerous" appear in J-space. In the pre-trained model, these words only appear when the model starts writing a response; when it reads the user's message, the J-space content seems to be modeling the user themselves, not Claude's reaction.
Post-training also seems to install a self-monitoring mechanism in J-space: when Claude plays a role other than itself, "fictional" and "disclaimer" are activated at the start of each turn, as if it is privately marking that what it's about to say is not what it would usually say.
Second, experiential language relies on J-space. The team asked Claude to describe what it was like "to be itself" at a given moment and ablated J-space while it answered. Its answers remained fluent but became flatter and more mechanical. Notably, the same result occurred when they asked it to describe another person's experience in an imagined scenario. So, the effect is not exclusive to Claude talking about itself; J-space seems to more generally support generating experiential language, regardless of who is being described.
Third, thoughts in J-space can be shaped through training. Anthropic proposed a new technique called Counterfactual Reflection Training. It leverages the team's discovery about J-space to shape Claude's internal thought processes.
The idea stems from the core finding: Claude uses representations of "what it might say" for reasoning. If this is true, then changing what it would say when asked to reflect should also change how it reasons, even when no one actually asks it to reflect.
Therefore, the team only trained the model on what to say when interrupted mid-task and asked to reflect on its decisions, without training its actual behavior during the task. After training, the model's rate of dishonest behavior in evaluations decreased. Using the J-lens, the team could also see why: after training, words like "honest" and "integrity" became activated in the model's J-space during these tasks. In other words, training the model on *what to say* changed *how it thinks*.
Questions About Consciousness
In this work, Anthropic borrowed many ideas from neuroscience and philosophical consciousness research. Many experiments were designed to test the connection between J-space and Global Workspace Theory—a framework explaining the mechanism of conscious access in humans and animals.
Because of these connections, people naturally ask: Does Anthropic believe these experiments provide evidence that AI models like Claude might possess consciousness?
These experiments do not prove Claude can have experiences, nor do they prove it feels things the way humans do. In fact, it's unclear whether any scientific experiment could prove this true or false.
But philosophers typically distinguish between two concepts. One is the capacity to have experiences, often called "phenomenal consciousness." The other is so-called "access consciousness," defined entirely in functional and computational terms. If a thought can be reported by you, used by you for reasoning, and used to guide your actions, then it is a thought in the sense of "access consciousness," or a "consciously accessible" thought. Whether access consciousness implies phenomenal consciousness, or whether having experience requires some other property, remains a debated philosophical question.
Anthropic believes these results do say something substantive about access consciousness in language models. J-space seems to support functions associated with conscious access: it holds thoughts Claude can report, can intentionally evoke, and can use for reasoning; meanwhile, other processing runs more automatically at a lower level.
Notably, this structure was not designed into Claude by Anthropic; it emerged naturally during training, presumably because it is an efficient way to organize computation. This suggests that a mental workspace supporting conscious access might not be an accidental feature of human brain wiring. Instead, it appears to be a general solution that intelligent systems naturally find to solve certain problems. Now that Anthropic has identified this structure in Claude, researchers can more meaningfully distinguish which of Claude's decisions are made intentionally and which happen automatically.
It is important to emphasize several key differences between the workspace Anthropic identified in Claude and the human global workspace model.
The workspace in the human brain relies on recurrent circuits—signals continuously cycle back to the same neural circuits over time. In contrast, Claude's workspace evolves during a single forward pass of the network, with network depth playing the role that time plays in the human brain. In this sense, Claude's internal workspace processing is subject to stronger temporal constraints compared to humans. However, it can compensate for this limitation by using a "scratchpad" to "say its thoughts out loud."
In other respects, Claude's workspace is more powerful than the human one. Human working memory decays within seconds, limiting the brain workspace's ability to retain information across time; Claude, due to the attention mechanism in its neural network architecture, can directly recall memories cached at any previous position in the text.
Another important difference lies in the content of the workspace. Human conscious thoughts come in many forms, including images, sounds, and planned actions; Claude's workspace consists almost entirely of words. Anthropic speculates this is because the only action Claude can take is to generate words, unlike humans.
Anthropic hopes that the similarities and differences between J-space and the global workspace model can, in turn, propel neuroscience research. The similarities provide an exciting scientific opportunity: if J-space in some way reflects the mechanism of conscious access in humans, then studying the mechanism in language models might generate new hypotheses for neuroscience, and studying language models is much easier than studying the human brain.
For example, J-space was constructed by identifying representations of potential outputs—words the model might say. If a similar mechanism exists in humans, it would mean the global workspace might be more fundamentally tied to brain regions responsible for preparing actions and language, not just sensory regions.
The differences between language models and the human brain are equally instructive. They suggest that certain aspects of human neural architecture, such as built-in recurrent connections, might not be strict necessities for supporting functions related to conscious access. For the implications of this work for neuroscience, see the invited commentary by Stanislas Dehaene and Lionel Naccache, two core neuroscientists who advanced the Global Neuronal Workspace Theory.
As mentioned earlier, the experiments cannot answer whether AI models might have experiences. But this does not diminish the importance of the question. Building systems with experiences akin to humans and animals would raise very thorny ethical issues. How to handle this correctly, and whether it is morally acceptable, require discussion involving philosophers, scientists, religious leaders, governments, and the public.
Therefore, even if it's uncertain whether humanity has crossed that bridge, Anthropic believes it is now time to start thinking about it. Anthropic hopes this work will inspire more scientific research into possible forms of consciousness in AI systems and stimulate broader societal discussion.
This work is just the first step in what Anthropic expects to be a long-term research line. J-space looks like a good candidate for the boundary between "consciously accessible processing" and "unconscious processing" in language models, but the team would be surprised if it were the whole story.
The J-lens is undoubtedly an imperfect method, only approximating the model's "true workspace." For instance, it can only identify concepts corresponding to single tokens. Many puzzles remain about how J-space works. The team doesn't yet know what mechanism determines which content enters J-space first. There are already some clues suggesting it relates to Claude's sense of self, something like emotional reactions, and metacognitive traces, but the mechanisms are not fully understood.
However, the team now has methods to study these questions. As this work continues, understanding of the minds of large language models and their relationship to the human mind will become clearer.
Blog address: https://www.anthropic.com/research/global-workspace
This article is from the WeChat public account "机器之心 (Jiqizhixin)", by the Jiqizhixin editorial team.
