Artificial intelligence company Anthropic has released new research suggesting that advanced large language models (LLMs) develop an internal mechanism resembling a "mental workspace" that helps them organize information and perform reasoning before generating responses.

The findings offer new insights into how AI systems process information internally but do not suggest that AI is conscious or possesses human-like awareness.

AI Appears to Use an Internal Reasoning System

According to Anthropic's study, modern language models maintain a small set of important internal representations that are separate from their broader computation.

Researchers describe this as a functional "mental workspace" where key pieces of information are gathered and organized before the model produces an answer. Rather than simply predicting the next word based on statistical patterns, the workspace appears to support more deliberate reasoning and structured decision-making.

Anthropic noted that this workspace represents only a small portion of the model's total computation, with most processing remaining automatic and distributed throughout the network.

Inspired by Human Neuroscience

The researchers compared this behavior to the Global Workspace Theory, a prominent neuroscience framework that explains how the human brain processes information.

According to the theory:

• The brain continuously processes vast amounts of information unconsciously.
• Only a limited subset enters a "global workspace," where it becomes available for conscious planning, reasoning, language, and decision-making.

Anthropic emphasizes that the comparison is functional rather than philosophical. While AI models exhibit information-processing patterns that resemble aspects of the theory, they are not conscious and do not possess subjective awareness.

Jacobian Lens Reveals Hidden AI Thought Processes

To investigate these internal mechanisms, Anthropic developed a new interpretability technique called the Jacobian Lens (J-Lens).

The tool enables researchers to observe concepts that the model is preparing internally before they appear in the final generated response.

Unlike traditional methods that analyze only the completed output, J-Lens provides a window into the intermediate reasoning steps, helping researchers understand:

• Which concepts the model considers.
• How information is organized.
• How internal representations evolve before an answer is produced.

This could significantly improve transparency in AI systems.

Why the Research Matters

The study contributes to the growing field of AI interpretability, which seeks to understand how complex machine learning models make decisions.

Greater visibility into internal reasoning can help:

• Improve AI reliability and safety.
• Detect hallucinations or flawed reasoning.
• Build more transparent and trustworthy AI systems.
• Enhance alignment between AI behavior and human intentions.

As AI becomes more capable, researchers argue that understanding how models reach conclusions may be just as important as evaluating the accuracy of the answers themselves.

Key Takeaway

Anthropic's research suggests that advanced AI models develop an internal "mental workspace" that organises information before producing responses, offering a closer look at how large language models reason. However, the study does not claim that AI is conscious. Instead, it highlights that AI can evolve computational structures that perform functions analogous to certain aspects of human cognition, underscoring the importance of interpretability as AI systems become increasingly sophisticated.