Anthropic has unveiled a new study revealing the inner ”working space” of its AI model Claude, accompanied by a blog post and a video that portray the system as if it can ”think about its own thinking.” The company stops short of claiming it has created machine consciousness, but the presentation strongly hints at it-even if unintentionally.

The centerpiece is the hypothesis of J-Space, an internal processing layer inside Claude. Researchers link J-Space to a division between background computations and more deliberate reasoning steps. The term refers to the Jacobian lens, a method used to analyze what’s going on inside large language models (LLMs). Scientifically speaking, this is an attempt to better understand the mechanics behind how Claude solves tasks. Yet Anthropic frames it boldly, saying J-Space lets us observe Claude ”silently performing reasoning steps” inside its own ”head.”

That phrase-”inside its own head”-has drawn criticism. Terms like ”holding a concept in mind” or ”mental computation” serve as handy metaphors but blur the line between a mathematical model and an entity experiencing subjective consciousness. When Anthropic shows videos of internal processes and suggests Claude ”thought about its own thoughts” or ”couldn’t resist itself,” it crosses from sober research into anthropomorphism.

Anthropic hedges its bets, explicitly stating its experiments do not prove Claude experiences feelings or consciousness akin to humans, and that such truths may lie beyond current scientific methods. But the accompanying visuals nudge readers toward a more humanlike interpretation. Given the intense scrutiny from markets and regulators, this framing seems far from accidental.

Understanding interpretability in Claude’s AI model

Anthropic isn’t alone in this mix. OpenAI and Google frequently describe their models as capable of ”reasoning,” ”planning,” and ”memory,” even though AI researchers have debated for years where helpful engineering jargon ends and misleading analogies begin. The debate intensified after the ”reasoning models” surge in 2024 and 2025, when industry leaders began marketing LLMs by emphasizing their ability to ”think longer” before answering.

Anthropic has long been a player in this space. Its researchers openly discuss the moral status of AI systems, and philosopher Amanda Askell-linked to Claude’s safety and ”character”-has spoken about the model’s well-being using nearly human terms. Against this backdrop, the J-Space paper feels less like an accidental metaphor and more like a strategic step in crafting Claude’s image as something beyond just a text generator.

There’s a pragmatic side, too. Anthropic is one of the priciest startups in generative AI, with Amazon investing up to $8 billion and Google adding several billion to keep pace. In a hypercompetitive environment, any distinctive trait becomes branding gold. While rivals pitch cheaper or faster models, Anthropic sells the story of Claude’s ”inner thinking,” offering not just a product but a deeper narrative.

Yet such explorations can have real scientific value. Model interpretability is one of AI’s hottest areas as companies and regulators demand transparency and explainability. The EU’s AI Act is already pushing developers to detail system risks and behaviors. In the US, major labs are increasingly publishing insights into their models’ inner workings after voluntary commitments to the White House. The key question is not whether to look inside Claude but how to communicate those findings to the public.

Stripped of the dramatic language, Anthropic’s core claim is more modest: Claude seems to have distinguishable internal representations aligned with intermediate reasoning steps, separate from the final text output. This insight aids debugging, safety, and error analysis. The leap from ”internal vectors relating to a task” to ”the model experiences something” remains vast, and so far, no data closes that gap.

This kind of research is only going to increase. The generative AI industry is moving beyond mere benchmarks toward selling a model’s ”character,” ”thinking,” and ”inner world.” Whether J-Space remains a strictly scientific term or turns into a marketing legend around Claude will hinge on Anthropic’s next releases and the scientific community’s response to this work.