Inside Claude: Anthropic finds AI uses a human-like reasoning workspace

Anthropic study finds Claude uses an internal reasoning workspace before generating responses, while stressing the findings do not suggest the AI is conscious

For years, the debate around artificial intelligence (AI) has centred on one question: Can machines ever think like humans? While AI systems can write essays, generate code and solve complex problems, they are still believed to lack qualities such as consciousness, self-awareness, emotions and genuine understanding. At the same time, rapid advances in AI have fuelled speculation about whether these systems could one day surpass human intelligence, making it increasingly important to understand what actually happens inside them before they generate a response.

 

While models such as Claude, ChatGPT and Gemini can perform increasingly complex tasks, the reasoning behind their responses remains largely hidden. As a result, researchers often describe them as "black boxes" because their outputs are visible, but their internal decision-making is difficult to explain, even for the companies that built them.

 

 

US-based AI entity Anthropic, in its research, tried to understand what happens inside its large language model (LLM) Claude before it generates a response. Instead of focusing on whether the model gives the right answer, the researchers studied how it solves problems internally. They found evidence of what they call a "global workspace", a temporary internal space where important information is gathered and processed before Claude produces its final response.

 

Anthropic makes it clear that the findings do not mean Claude is conscious or has human-like awareness. Instead, the research suggests that advanced AI models may have developed internal reasoning mechanisms similar to those described in theories of human cognition. The company believes this could help researchers better understand how AI works and improve its transparency and safety.

 

From neuroscience to artificial intelligence

 

Anthropic says the idea behind the research comes from neuroscience, particularly how scientists study conscious and unconscious processing in humans. Much of the brain's activity happens automatically, without people being aware of it. Only a small amount of information enters conscious awareness, where it can be used for reasoning, planning and problem-solving. The researchers wanted to examine whether a similar distinction exists inside Claude.

 

To explore this, Anthropic turned to “Global Workspace Theory”, a framework that explains how information becomes consciously accessible. According to the theory, the brain selects a small amount of important information and makes it available to different parts of the brain for reasoning and problem-solving. The researchers investigated whether Claude has a comparable internal workspace that supports these functions.

 

Anthropic wanted to investigate whether something comparable emerges inside a LLM. Importantly, the researchers were not trying to prove that AI has consciousness. Instead, they wanted to determine whether AI develops organisational structures that resemble some of the computational mechanisms described in human cognition.

 

Their experiments identified what Anthropic calls the ‘J-space’, an internal reasoning space where Claude appears to gather and process information before generating a response.

 

What is the 'J-space'

 

To explore Claude's internal activity, Anthropic developed techniques to monitor patterns inside the neural network while the model worked through different tasks. Rather than identifying individual thoughts, the researchers found recurring patterns that appeared to represent concepts the model was actively using. They collectively referred to this internal representation as J-space. Anthropic says the J-space was not programmed into Claude but emerged naturally during the model's training.

 

J-space can be thought of as an internal working area where Claude temporarily stores ideas while solving a problem. These ideas are not always visible in the model's final response. Instead, they exist behind the scenes and influence how Claude arrives at its answer.

 

Unlike a conversation with a user, where only the finished response appears on screen, J-space captures information that remains hidden during normal interactions. According to Anthropic, this makes it possible to observe parts of the model's reasoning process that would otherwise remain inaccessible.

 

The researchers argue that this workspace behaves less like a collection of stored facts and more like an active reasoning environment where information is combined, updated and manipulated before the final response is generated.

 

What the experiments revealed

 

To understand whether J-space played a meaningful role, Anthropic conducted a series of experiments. One involved mathematical reasoning. Claude was asked to solve arithmetic problems while providing only the final answer. Although the visible response contained no intermediate calculations, the researchers observed changing internal representations that corresponded to successive steps in the calculation.

 

This suggests that the model was not jumping directly to the final answer. Instead, it appeared to work through intermediate reasoning internally before generating its response.

 

Another experiment examined whether Claude could deliberately maintain an internal idea while performing a different task. Researchers instructed the model to think about a specific object while simultaneously copying unrelated text. Although the visible output focused entirely on the copying task, the internal workspace continued to represent concepts associated with the requested object.

 

This finding suggests that Claude can maintain internal representations independently of what it is writing externally. In other words, the model appears capable of separating its internal reasoning from its outward response, at least in certain situations.

 

Anthropic also tested whether this internal workspace could be suppressed. Even when instructed not to think about a particular concept, traces of that concept continued appearing in the internal representations. This resembles the difficulty humans often experience when trying not to think about something.

 

Other key findings

• J-space evolves with training: Anthropic found that J-space exists before post-training but develops a distinct "Claude" perspective after the model is trained to act as an AI assistant.
• It detects risks internally: In one test, words like "warning" and "dangerous" appeared in J-space before Claude responded to a potentially harmful user query.
• It supports self-monitoring: During roleplay, J-space activated words such as "fictional" and "disclaimer", suggesting Claude internally tracks when it is acting as a character.
• It helps generate natural responses: Disabling J-space made Claude's responses more mechanical, indicating the workspace supports richer, experience-based language.

Why switching off the workspace mattered

 

One of the most significant experiments involved reducing access to the internal workspace while leaving the rest of Claude's neural network intact. The results revealed an important distinction between language generation and reasoning.

 

Without this workspace, Claude could still perform relatively straightforward tasks. It remained capable of writing fluent text, answering simple factual questions and producing grammatically correct responses.

 

However, when the tasks required multiple reasoning steps or the integration of different pieces of information, performance declined noticeably. According to Anthropic, this suggests the internal workspace is particularly important for higher-order reasoning rather than basic language generation. It appears to help the model organise information before producing more complex responses.

 

Although the experiments do not explain every aspect of Claude's behaviour, they indicate that advanced reasoning may rely on specialised internal mechanisms rather than emerging solely from next-word prediction.

 

Does this mean Claude is conscious

 

The company says its research does not show that Claude has feelings, subjective experiences or human-like consciousness. Instead, the study focuses on what researchers call "access consciousness," a concept from neuroscience that refers to information a system can access, reason with and use to make decisions.

 

Anthropic argues that the J-space appears to perform many of these functions. It stores information that Claude can deliberately use while solving problems, while much of the model's other processing continues automatically in the background. However, the researchers emphasise that this should not be confused with phenomenal consciousness, which refers to the ability to have subjective experiences or emotions.

 

The company says there is currently no scientific method to determine whether an AI system possesses that kind of awareness. One of the study's key findings is that J-space was not intentionally designed by Anthropic. Instead, it emerged naturally during Claude's training, suggesting that advanced AI models can develop internal reasoning mechanisms on their own.

 

At the same time, Anthropic notes that Claude's internal workspace is fundamentally different from the human brain. While human conscious thought combines images, sounds, memories and actions, Claude's workspace is almost entirely built around language because words are the model's primary way of interacting with the world. The company also points out that human working memory fades over time, whereas Claude can retrieve information from earlier parts of a conversation using its attention mechanism.

 

Rather than claiming AI is conscious, Anthropic says the research offers a new way to understand how language models reason internally. The company believes this could improve AI transparency and safety, while adding that questions around machine consciousness remain unresolved and require broader scientific and ethical debate. 

 

What does this mean for the future of AI

 

Anthropic believes interpretability research will become increasingly important as AI models grow more capable. Current frontier models already influence software development, customer service, scientific research and enterprise decision-making. Yet their internal processes remain difficult to explain.

 

Improving transparency could help developers identify hidden biases, reduce hallucinations, strengthen safety testing and increase confidence in AI-assisted decisions. For regulators, such research may also support future governance efforts by providing more reliable methods for auditing advanced AI systems. The study reflects a broader shift within AI research, from measuring what models can do to understanding how they do it.