If you've watched the 2004 sci-fi film I, Robot, you'll likely remember the story of robots developing increasingly human-like intelligence and reasoning abilities. While Anthropic is not claiming that its Claude AI model has become conscious, the company's latest research draws an intriguing comparison with neuroscience after identifying what it calls a "J-space", an internal computational workspace that resembles the global workspace theory of the human brain. According to the company's paper, the structure emerged during training and helps Claude organise information while solving complex reasoning tasks.According to the research, the structure was not intentionally built into Claude's architecture but instead emerged during training. Anthropic says the finding may offer new insights into how advanced AI systems organise computation when handling difficult reasoning tasks.The research also introduces a new analytical tool called J-Lens, or Jacobian Lens, which the company developed to examine Claude's internal computations and better understand how the AI model reaches its responses.Research draws comparison with global workspace theoryAnthropic's findings are based on the global workspace theory, a concept in neuroscience that suggests only a small portion of the brain's activity becomes consciously accessible at any given time.According to the theory described in the paper, while much of the brain continuously performs automatic functions such as regulating breathing and processing vision, certain thoughts enter a "global workspace" where they can be held in mind, reasoned about and shared across different parts of the brain.The researchers say they observed a comparable computational structure inside Claude, which they refer to as J-space.To explain the concept, the paper compares Claude's internal processes to a large corporate office. While numerous background operations occur simultaneously, a central boardroom - the J-space- serves as the place where important concepts are gathered when the model needs to solve a complicated problem, work through difficult situations or explain the reasoning behind its responses.According to Anthropic, this workspace becomes active when Claude needs to coordinate information across different parts of the model.J-lens developed to study Claude's internal reasoningTo investigate how Claude organises information internally, Anthropic researchers developed J-lens, a mathematical tool designed to filter out background computations and focus on the concepts currently influencing the model's output.According to the paper, J-lens examines Claude's internal calculations and predicts how specific signals affect the likelihood of particular words or responses appearing later in the model's output.The researchers explain that the tool can answer questions such as how a signal appearing at a particular stage of the model's computation changes the probability of a specific word being generated several words later.Anthropic says the tool provides researchers with a way to examine the model's reasoning process rather than focusing only on its final responses.Researchers carried out multiple experimentsAccording to the paper, Anthropic researchers conducted several experiments to determine whether the proposed J-space existed within Claude.In one experiment, researchers asked Claude what it was thinking about while simultaneously using J-lens to observe the concepts present within the J-space.Another experiment instructed Claude to keep the concept of fairness in mind. According to the researchers, the model intentionally brought the concept into the J-space and maintained it there while reasoning through the task.The paper also describes experiments involving chess problems. Researchers found that the J-space functioned like a scratchpad during complex reasoning. When they altered the contents of the workspace, Claude's final output also changed.In another observation, Anthropic reported that once a concept entered the J-space, other parts of the model performing unrelated tasks were still able to access and use that information.According to the researchers, these observations suggest the workspace acts as a shared computational area that coordinates information across different parts of the model.Anthropic says the structure emerged during trainingOne of the key findings highlighted in the paper is that the J-space was not explicitly designed into Claude's architecture.According to Anthropic, the structure emerged naturally during training because it provided what researchers describe as a useful way of organising computation.The company compares this to the idea of convergent evolution in biology, where unrelated species independently develop similar characteristics because they represent effective solutions to common challenges.Anthropic suggests that the emergence of a workspace resembling the global workspace theory may indicate that such computational structures are an efficient way for systems to process complex reasoning tasks.However, the research focuses on understanding Claude's internal computational processes and does not state that the AI model possesses human consciousness or a "soul." Instead, the paper presents the findings as an observation about how advanced AI systems may organise information while performing complex reasoning.Why the findings matterAccording to Anthropic, tools such as J-Lens could help researchers better understand how large language models make decisions internally rather than treating them as black boxes.By examining the concepts active during reasoning and observing how changes affect the model's responses, researchers hope to gain deeper insight into how AI systems organise computation as they solve increasingly complex tasks.The company says the findings also open new avenues for studying the internal behaviour of AI models, potentially helping improve transparency and interpretability as these systems continue to evolve.Frequently Asked QuestionsQ1. What is the J-space described in Anthropic's research?According to Anthropic, J-space is an internal computational workspace within Claude that becomes active when the AI model needs to reason through complex problems, explain its logic or hold important concepts in mind.Q2. What is J-lens?J-lens, or Jacobian Lens, is a mathematical tool developed by Anthropic to study Claude's internal computations and examine how different concepts influence the model's responses.Q3. Why do researchers compare J-space with the human brain?The paper says the structure resembles the global workspace theory in neuroscience, which describes how certain thoughts become accessible across different parts of the brain during reasoning.Q4. Did Anthropic intentionally build J-space into Claude?No. According to the research, the J-space was not part of Claude's planned architecture and instead emerged during the model's training process.Q5. Does the research claim Claude is conscious?No. The paper discusses similarities between Claude's internal computational structure and the global workspace theory but does not state that the AI model possesses human consciousness or a soul.end of article