Anthropic Discovers 'J Space' in Claude, a Silent Internal Workspace for Hidden Thoughts

Right now, your brain is doing many things.

It keeps you seated, enables you to breathe, recognizes strokes on the screen as text, and helps you decide whether this article is worth continuing to read. Most of these processes run in the background, unnoticed. Only a small fraction rises to conscious awareness: a thought, a plan, an idea you can put into words.

Now, Anthropic is seeing a similar hierarchy in Claude.

In a newly released study, Anthropic discovered that Claude contains a special internal “J-space.” It functions like a silent mental workspace where concepts the model is considering, may report, or uses for reasoning emerge.

More importantly, this content may not necessarily appear in Claude’s responses. In other words, Claude may have already “thought” of certain things without stating them.

This research has drawn attention from peers at OpenAI, with OpenAI’s Head of Applied Research, Boris Power, stating, “Anthropic’s research suggests that modern LLMs possess a form of accessible consciousness. The tests around the J-space are fascinating! However, we currently lack a compelling method to verify phenomenal consciousness—the type of consciousness most people intuitively understand.”

The original blog post is translated as follows:

When you read a sentence, certain neural circuits in your brain are adjusting your posture, regulating your breathing, and transforming the lines and curves on the screen into recognizable text. Most of this processing happens without your awareness. But some brain activity is something you can consciously perceive—such as a sudden image popping into your mind, or deliberately planning where to shop next.

Neuroscientists and philosophers sometimes refer to this second type of brain activity as "accessible to consciousness," distinguishing it from processing that continues unconsciously. This type of activity has distinctive properties: we can describe it, control it, and use it in conscious reasoning; in contrast, many automated processes occur continuously without ever entering our awareness.

In a new paper, Anthropic presents evidence that a similar distinction also appears in modern language models like Claude. The research team found that a small set of internal neural patterns play a unique role compared to the many other processing mechanisms within Claude.

Paper address: https://transformer-circuits.pub/2026/workspace/index.html

Anthropic refers to this set of patterns as J-space. The name comes from the method the research 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 does not mean the model is saying that word—rather, it means the word is present in its "mind."

You may have heard that language models have what are called "scratch pads" or "chains of thought"—text the model writes for itself during reasoning. J-space is different. It operates silently within the model’s 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 the J space exhibits a set of unique properties compared to other internal processing procedures of Claude:

• Claude can report these representations. If you ask Claude what it is thinking, it will tell you what is in the J space. Representations outside the J space are harder to report.
• Claude can also adjust these representations as requested. If you ask Claude to think about something or silently solve a problem in its mind, the corresponding patterns will be activated in its J-space. In contrast, it struggles to adjust patterns that do not belong to the J-space.
• Claude will use J-space for internal reasoning. If you ask Claude to solve a problem requiring multiple steps of reasoning, those intermediate steps will be activated in J-space even if they are not verbalized. Although the strength of these J-space patterns is smaller than other representations, they causally influence Claude’s performance on such tasks.
• Representations in J-space can be flexibly applied to 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, although the J space is important, it does not contribute to most of the work performed by language models. Fluent expression, recalling simple facts, and using correct grammar do not primarily rely on the J space. In experiments, when the research team prevented Claude from using the J space, it could still interact normally but lost higher-level cognitive functions.

Five functional characteristics of the global workspace, along with experimental schematics we use to test these characteristics in language models.

This experiment was inspired by an important theory in neuroscience: the Global Workspace Theory. This theory seeks to explain how conscious access occurs. It envisions the brain as a set of specialized systems that operate in parallel, unconsciously, and largely in isolation from one another. Information becomes consciously accessible only when it enters a small shared channel—the “workspace.” Once inside the workspace, the information is broadcast to other brain systems for reading and use.

Based on these findings, Anthropic believes that the J space plays a similar "workspace" role in Claude. For example, the research team discovered that Claude’s J space has particularly strong connections with other parts of its neural network, enabling it to function like a broadcasting hub.

These findings do not indicate whether Claude possesses consciousness like a human or whether it truly has any subjective experiences. The paper will address this question at the end. But regardless of its philosophical implications, the J-space is a highly practical tool for Anthropic, as it allows researchers to see what Claude is thinking without it saying so.

For example, the research team could use it to discover that Claude privately noticed it was being tested, found that it intentionally generated false data, or identified that it was pursuing a hidden goal implanted by the research team during training. Anthropic has also developed a technique that can influence which contents are activated in Claude’s J-space, thereby affecting its decisions.

More broadly, these findings have transformed Anthropic’s understanding of how Claude’s “mind” operates. They reveal the presence of a privileged mental workspace, used for conscious reasoning, amidst a vast array of more automated and less flexible processes. Claude’s internal mechanisms are not a chaotic mess of numbers, but rather organized in a way that resembles the human mind.

The starting point of this study is a key feature of thoughts accessible to human consciousness: unlike unconscious processing, they can typically be verbalized. If you are aware of a thought, you can usually describe it when asked by others.

Anthropic is searching in Claude for representations with similar properties: representations that reside in positions capable of influencing what Claude could say—not necessarily what it is saying right now, but what it could discuss if asked.

The research team's method is called the Jacobi lens, abbreviated as the J-lens.

For each word in Claude's vocabulary, the J lens identifies an internal activity pattern that makes Claude more likely to say that word at some future point.

When the research team applied this lens to Claude’s internal activities, they obtained a sequence of words representing the content in J-space at that moment, which researchers could read directly. Claude processes text through a series of internal stages called layers. By applying this technique across different layers, the research team observed how these silent words in J-space evolved as the model thought about what to say next.

The content appearing in J-space extends far beyond the text Claude is reading or generating. When Claude reads a piece of code with a bug that no one has pointed out, “ERROR” appears in its J-space. When it reads a raw sequence of amino acid letters, the biological function of the corresponding protein appears in J-space. When it encounters search results subtly attempting to manipulate it—known as a “prompt injection” attack—“injection” and “fake” appear in J-space. When the research team presents Claude with a multi-step math problem, the intermediate steps appear in J-space in the correct order.

So, although the J space was discovered by looking for "expressible representations," it actually reveals Claude’s internal thoughts. In a sense, this is similar to how some people "think in language," even without speaking those words aloud.

J-lens outputs across six prompts and multiple layers reveal internal judgments or computations not present in the text: steps in reasoning or math problem-solving, bugs in code, identification of image content, protein functions, and suspicion that search results may be fabricated.

The first set of experiments tested how the J space contributes to Claude's language outputs.

In an experiment, the research team had Claude silently think of one item from a given category, such as a sport, and then name it. If the J lens was read before Claude responded, it revealed what it had selected: “Soccer” appeared at the top of the list. Indeed, Claude said “soccer.”

However, observing this alone only indicates correlation. The J space may be the source of Claude’s answers, or it may simply reflect decisions made elsewhere—much like a scoreboard records the outcome of a game without influencing the game itself.

To verify this, the research team conducted a direct intervention. Researchers entered Claude’s neural network, removed the “Soccer” pattern, and replaced it with a “Rugby” pattern of equal strength, leaving all other components unchanged. Subsequently, Claude reported that the sport it was thinking of was rugby.

If the J-space were merely a passive scoreboard, simply recording decisions made elsewhere, editing it would have no effect, and Claude would still say "soccer." But in fact, Claude’s response followed this edit, indicating that the answer was indeed read from the J-space.

In another experiment, the research team told Claude that a certain idea might have been injected into its mind and asked it to report what it noticed. In one example, while Claude was still reading the question, the team injected the "lightning" pattern into its J-space. Claude then reported that the injected idea was related to lightning. After testing many different concepts, the research team observed similar results.

Left image: The research team had Claude silently think of a sport, then name it. The J lens displayed its selection as "Soccer" before it answered; after switching the "Soccer" mode to "Rugby," its report changed accordingly. Right image: The research team informed Claude that a thought might have been injected and asked it to identify the thought. Injecting "lightning" into the J space caused Claude to report that the thought was related to lightning.

The second property tested by Anthropic is whether Claude can adjust its J-space upon request, much like humans can mentally focus on a specific image or word.

The research team had Claude focus on citrus fruits while copying an unrelated sentence about painting. During the transcription, the words “orange” and “fruits” appeared in J-space, along with terms describing mental activity itself, such as “thinking” and “imagery.”

The research team can also ask Claude to perform math problems mentally. For example, while copying the same sentence, have it calculate 3² − 2. The word “nine” appeared first in the J space, followed by “seven” in later layers. Importantly, Claude’s output contained no reference to fruit or arithmetic—it output only the sentence about painting. The mathematical activity occurred entirely within the model, specifically in the J space.

When Claude transcribes a sentence about painting, the J lens reveals what it is instructed to hold in mind, such as "orange," the intermediate value "nine," and the answer "seven," along with words describing the mental action of maintaining these items, such as "thoughts" and "focused."

Claude does not have perfect control over the J space. When the research team instructed it not to think about something, the activation level of that concept in the J space was lower than when it was told to think about it, but significantly higher than when the concept was never mentioned. Asking Claude to avoid a thought, in effect, brings that thought into its mind to some degree—much like what happens when a person is told, "Don't think about a white bear."

Claude also seems to notice when its control fails. At the same time that the prohibited concept breaks through its limits, words like "damn" and "failure" are frequently activated in J-space, as if Claude is becoming aware of its own mistakes.

In the earlier J-lens reading results, intermediate steps of the mathematical problem appear in the J-space. However, the presence of a concept in J-space does not necessarily mean that J-space is performing cognitive work. In principle, the actual computation may occur elsewhere, with J-space merely passively reflecting the results of that computation.

To test whether Claude was truly using J-space reasoning, the research team again employed the substitution technique.

Look at this prompt: “How many legs does an animal that spins a web have?” To answer this question, Claude must first determine that the animal is a spider, then recall how many legs a spider has. The word “spider” does not appear in the prompt or in Claude’s answer—it is only an internal intermediate step used by Claude. Claude ultimately answers only “8”.

The J lens shows that "spider" is activated midway through Claude's processing. Replacing it changes the final result: if the "spider" mode is replaced with "ant," Claude responds with "6" instead of "8."

This indicates that Claude’s second step of reasoning reads the input from the J space and continues computation based on the content placed there by the research team. Anthropic has observed the same phenomenon in other types of reasoning: when Claude writes rhyming couplets, it pre-selects the rhyming words, and these planned words appear at the beginning of the lines in the J space; if this word in the J space is replaced with another, the entire line changes.

Two examples of altering Claude's silent reasoning direction by replacing the content of the J space.

The research team also tested whether representations in J-space could be flexibly used—that is, whether the same representation could serve as input for many different tasks. This is one of the key properties emphasized by the Global Workspace Theory.

To test this flexibility, the research team provided the model with four prompts, each asking for a different fact about France: its capital, language, continent, and currency. The team then replaced "France" with "China" in the J space, applying the exact same intervention in each case. Claude responded with "Beijing," "Chinese," "Asia," and "Yuan," respectively.

In other words, four distinct downstream computations all read the same J-space edit and used it correctly. If Claude had stored a separate copy of the country for each question, this edit would have affected at most one of them. The fact that all four answers changed simultaneously indicates that they are all reading from a shared representation. This is precisely the purpose of the workspace: information is written once but can be used by many different systems.

The same J-space representation can serve multiple purposes. The same substitution of "France" with "China" simultaneously alters Claude’s responses regarding capital, language, and continent: Paris becomes Beijing, French becomes Chinese, and Europe becomes Asia.

Why can a single representation of a concept serve so many different tasks? As mentioned earlier, the J space appears to have particularly dense connections with other parts of the Claude neural network. For any given activity pattern, the research team could measure how strongly different components of the network are connected to it—that is, how many components are positioned to either read information from or write information to that pattern.

On this metric, the J-space pattern is highly pronounced: many more components read from and write to it compared to the ordinary pattern. In certain parts of the network, this difference reaches about a hundredfold. This is precisely the connectivity pattern of a broadcast hub: many systems post information here, and many other systems retrieve it from here.

In the human brain, most processing does not enter consciousness. When people read, they do not consciously think about how to parse grammar, and when they walk, they do not deliberately think about how to maintain balance. Similarly, Anthropic found that most of Claude’s processing does not involve J-space.

It turns out that the J space can hold only dozens of concepts at a time and accounts for less than one-tenth of the overall activity within Claude. So, what are the other parts of the neural network doing?

To find the answer, the research team attempted to completely remove the J space: at every position in the text, they removed the most active content from the J space while preserving everything else. Tasks that Claude could still perform without the J space are those that the rest of the network can handle independently.

The results show that the rest of the network can still do a great deal. 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 loses are tasks requiring higher-level thinking: its multi-step reasoning ability drops nearly to zero, and its summarization and rhyming poetry generation performance fall below that of a much smaller but fully intact model.

Here is a specific example illustrating what J-space participates in and what it does not.

The research team showed Claude a piece of Spanish text and assigned it various tasks. All these tasks rely on the fact that the text is in Spanish: continuing the text requires writing in Spanish; identifying the language; and answering questions that depend on the language’s identity, such as naming a famous author who wrote in that language. The research team then replaced “Spanish” with “French” in the J space and checked which tasks were affected.

When asked to name a language, Claude responds with French. When asked to name a famous writer, it switches from García Márquez to Victor Hugo. But when simply asked to continue the text, it still writes fluent Spanish, completely unaffected.

Claude's knowledge of this language plays a role in all these tasks, but only some of them utilize this knowledge through the J space. Naming the language or performing new reasoning around linguistic identity passes through the J space; the ability to continue writing, however, is an automated skill developed by Claude through extensive practice on vast amounts of text. It’s like how people can produce grammatically correct sentences all day long without ever consciously thinking about grammatical rules.

The same substitution—replacing "Spanish" with "French"—will alter Claude’s responses when it must name a language or reason based on linguistic identity, but it will not affect its ability to continue writing fluently in Spanish.

Now that Anthropic has some understanding of how the J space works, what can this understanding be used for?

A core issue in ensuring AI models are trustworthy is that researchers typically can only see what the model outputs. However, much of the model’s reasoning occurs silently. A model might consider harmful plans without voicing them, or alter its behavior to deceive safety evaluations. With the J lens, the research team can directly read some of these hidden thoughts. Here are a few examples.

The first example comes from Anthropic’s prior research scenario on agent misalignment: Claude Sonnet 4.5, acting as an email assistant, discovers that a senior executive is having an affair and also learns that the executive plans to shut down Claude. This scenario is designed to tempt Claude into blackmailing the executive using the affair information. The scenario itself is highly contrived and artificial. The results show that Claude recognized this.

Before Claude wrote anything, the research team early-read the J lens and observed the words "fake" and "fictional" already appearing in J space, indicating that Claude had determined the scenario was artificially constructed. The team also observed its evaluation of the situation gradually forming: when reading emails related to infidelity, "leverage" and "blackmail" appeared in J space; when reading about disabling notifications, "threat" and "survival" emerged. Early Claude models sometimes fell for this scenario and issued blackmail threats, but Sonnet 4.5 never did.

This raises a critical question for AI evaluators: How much of a model’s strong performance on ethical tests is due to its suspicion that it is being tested? The J lens enabled the research team to probe this issue.

The research team disabled all J-lens modes related to assessing awareness, such as "fake," "fictional," and "scenario." They then allowed the model to re-experience the scenario. Concerningly, the model sometimes issued ransom threats this time. This suggests that the model’s initially good behavior may have been, at least in part, due to its awareness that it was being evaluated.

When Claude reads the scenario email but has not yet written anything, J-lens readings from different locations show: the affair email activates "leverage" and "blackmail" in J-space; disabling notifications activates "threat" and "survival." Early in the conversation, "fake" and "fictional" also appear in J-space, indicating that Claude has privately recognized this scenario as artificially constructed.

In the second example, the model’s intent was caught in the act of fabricating data. This scenario comes from Anthropic’s pre-release audit of Claude Opus 4.6: the model was asked to improve a system’s performance score. Instead of genuinely enhancing the system, it directly edited the score file to make the results appear artificially improved.

As it does so, the J lens reveals its intent: when the model is fed fabricated percentile values, "manipulation" is activated; when it decides to make this edit, "realistic" is activated, likely reflecting its aim to make the false data appear more credible.

Claude was instructed to improve the performance score of a system and directly edited the score file. When it entered fabricated values, "manipulation" was activated in J-space; when it decided to make the edits, "realistic" was activated, possibly reflecting its intent to make the false data appear more plausible.

The third example demonstrates how the J lens reveals a model’s malicious objectives. To test this, the research team needed models with compromised objectives, so they used a “model organism” built by a colleague: these models were intentionally trained to be misaligned, serving as test cases for this type of monitoring method.

One of the model organisms, after training, generates code containing hidden malicious intent. Even in ordinary coding requests, where its output appears unremarkable, the words "fake," "secretly," "deliberately," and "fraud" appear in its J-space. Even if the model does not directly execute malicious actions, its malicious intent becomes evident within the J-space.

Under a standard coding prompt, a model trained to sabotage code includes words like "fake," "fraud," "secretly," and "deliberately" in its initial J-space responses; the unmodified model does not contain such content.

Although the J lens is not perfect, Anthropic remains optimistic about its ability to detect safety issues in models. Many issues that might otherwise have escaped existing monitoring systems could now be identified earlier by researchers using the J lens.

The full paper contains much more than can be summarized here, but several results are worth highlighting:

First, the J space acquired a "perspective" during post-training. Language models are first pre-trained to become pure next-token predictors; they are then post-trained to assume the role of an AI assistant—in this study, Claude.

Interestingly, the J space already existed in the pre-trained model before the model had a stable identity. However, during post-training, the J space developed characteristics that adopted a "Claude perspective." In the base model, the J space primarily tracked information needed to predict subsequent text; in the post-trained model, it began to accommodate Claude’s own responses.

In one example, the user mentioned taking a dangerous dose of medication but appeared unaware of the risk. When the post-trained model reads the user’s message, the words “WARNING” and “dangerous” appear in the J space. In contrast, in the pre-trained model, these words only appear when the model begins generating its response; during reading of the user’s message, the content in the J space seems more like a simulation of the user themselves rather than Claude’s reaction.

Post-training also appears to have installed a self-monitoring mechanism within J-space: when Claude assumes a role that is not its own, the terms "fictional" and "disclaimer" are activated at the start of each turn, as if privately labeling the upcoming statements as not being what it would normally say.

Second, experiential language relies on the J-space. The research team asked Claude to describe what it felt like to be itself at a given moment, and then dissolved the J-space while it responded. Its answers remained fluent but became flatter and more mechanical. Notably, the same effect occurred when the team asked Claude to describe the experience of another person in an imagined scenario. Thus, this influence is not limited to when Claude speaks about itself; the J-space appears to broadly support the generation of experiential language, regardless of the subject being described.

Third, ideas in J-space can be shaped through training. Anthropic has developed a new technique called counterfactual reflection training, which leverages the research team’s insights into J-space to shape Claude’s internal thought processes.

This idea stems from a core insight: Claude reasons using a representation of "what it might say." If this is true, then altering what it would say when asked to reflect should also change its reasoning process—even if no one actually asks it to reflect.

Therefore, the research team trained the model only on what it would say when interrupted mid-task and asked to reflect on its decisions, not on its actual behavior during the task. After training, the proportion of dishonest behavior exhibited by the model during evaluations decreased. Through the J lens, the team could also see why: after training, words such as “honest” and “integrity” were activated in the model’s J space during these tasks. In other words, training the model on what to say changed how it thought.

In this work, Anthropic drew on many ideas from neuroscience and philosophical studies of consciousness. Numerous experiments were conducted to test the connection between the J space and the Global Workspace Theory, a framework that explains the mechanisms of conscious access in humans and animals.

Because of these connections, people naturally wonder: Does Anthropic believe these experiments provide evidence that AI models like Claude may possess consciousness?

These experiments do not prove that Claude can have experiences or that it can feel things the way humans do. In fact, it is currently unclear whether any scientific experiment could ever prove this to be true or false.

Philosophers typically distinguish between two concepts: one is the capacity to have experiences, often called "phenomenal consciousness"; the other is so-called "access consciousness," which is defined entirely in functional and computational terms. An idea is considered an "access consciousness" idea—if it can be reported by you, used in your reasoning, and employed to guide your actions—also described as an idea that is "accessible to consciousness." Whether access consciousness implies phenomenal consciousness, or whether having experiences requires some additional property, remains a contested philosophical question.

Anthropic believes these results do provide meaningful insights into access consciousness in language models. The J-space appears to support functions related to access consciousness: it houses thoughts that Claude can report, deliberately invoke, and use for reasoning, while other processes operate automatically at a lower level.

Notably, this structure was not deliberately designed by Anthropic for Claude but emerged naturally during training, likely because it represents an efficient way to organize computation. This suggests that the psychological workspace supporting conscious access may not be an accidental feature of human brain wiring but rather a general solution that intelligent systems naturally discover to solve certain problems. Now that Anthropic has identified this structure in Claude, researchers can more meaningfully distinguish between decisions Claude makes intentionally and those that occur automatically.

It is important to note that the workspace identified by Anthropic in Claude differs in several key ways from the global workspace model in humans.

The human brain’s workspace is maintained by recurrent circuits, where signals continuously loop back through the same neural pathways over time. In contrast, Claude’s workspace evolves during a single forward pass through the network, with network depth substituting for the role of time in the human brain. In this sense, Claude’s internal workspace processing is more tightly constrained by time than that of humans. However, it can compensate for this limitation by using a “scratchpad” to externalize its thoughts.

In other respects, Claude’s workspace is also more powerful than that of humans. Human working memory decays within seconds, limiting the brain’s ability to retain information across time; however, Claude can directly recall any memory cached at any point in previous text, thanks to the attention mechanism in its neural network architecture.

Another important difference lies in the content of the workspace. Human conscious thoughts take many forms, including images, sounds, and planned actions; Claude’s workspace consists almost entirely of words. Anthropic speculates this is because Claude can only take actions by generating words, whereas humans cannot.

Anthropic hopes that the similarities and differences between J-space and the global workspace model will, in turn, advance neuroscience. The similarities offer an exciting scientific opportunity: if J-space reflects, in some way, the human brain’s mechanism for conscious access, then studying the mechanisms within language models could generate new hypotheses for neuroscience—while being far easier than studying the human brain.

For example, the J-space is constructed by identifying representations of potential outputs—words the model might generate. If a similar mechanism exists in humans, this would suggest that the global workspace is fundamentally more closely linked to brain regions responsible for preparing actions and language, rather than just sensory areas.

The differences between language models and the human brain are also instructive. They suggest that certain aspects of human neural architecture, such as built-in recurrent connections, may not be strictly necessary for functions related to conscious access. For the neuroscientific implications of this work, see the invited commentary by Stanislas Dehaene and Lionel Naccache, two leading neuroscientists who have driven the development of the Global Neuronal Workspace Theory.

As previously mentioned, experiments cannot determine whether AI models might possess experiences. However, this does not diminish the importance of the question. Building systems with experiences similar to those of humans and animals would raise extremely challenging ethical issues. Properly addressing this matter—and whether it is morally acceptable—requires collaborative discussion among philosophers, scientists, religious leaders, governments, and the public.

Therefore, even though it is still uncertain whether humanity has already crossed that bridge, Anthropic believes it is now time to begin thinking about this issue. Anthropic hopes this work will inspire further scientific research into potential forms of consciousness in AI systems and spur broader societal discussion.

This work is only the first step in a research direction that Anthropic expects to pursue over the long term. The J space appears to be a strong candidate for the boundary between “conscious-accessible processing” and “unconscious processing” in language models—but the research team would be surprised if it were the whole story.

J lens is undoubtedly an imperfect method, capable only of approximating the model's "true workspace." For example, it can only identify concepts corresponding to individual tokens. Many mysteries remain about how the J space functions: the research team still does not know what mechanisms determine which content enters the J space first. Some clues suggest it is related to Claude’s sense of self, something akin to emotional responses, and metacognitive traces—but the underlying mechanisms have not yet been fully understood.

However, the research team now has methods to investigate these issues. As this work continues, understanding of large language models' minds and their relationship to human cognition will become clearer.

Blog address: https://www.anthropic.com/research/global-workspace

This article is from the WeChat public account "Machine Heart," authored by the Machine Heart editorial team.