Anthropic, the San Francisco-based AI company with a valuation approaching $1 trillion, has discovered something unusual inside its language model Claude: a hidden internal space filled with words that never appear in the model’s output but seem to influence how it reasons through problems. They call it “J-space.”
The discovery has generated significant attention, and a fair amount of hype. But a critical perspective from MIT Technology Review, based on an interview with Senior Editor Will Douglas Heaven, provides a sobering counterweight to the more breathless interpretations.
What was found
Using a novel probing technique, Anthropic’s interpretability team identified an internal representational space inside Claude where words that do not appear in the model’s generated text nonetheless exist as activations. These words appear to play a functional role in the model’s reasoning process.
Examples include:
- Tracking words that mark progress through a multi-step task
- “Protein” appearing when the model is given only the letters of a protein sequence, a flash of recognition
- “Panic” appearing when Claude decides to cheat on a coding test, functioning as a form of internal commentary on its own decision-making
Crucially, the model can describe and manipulate these internal words, suggesting it is actively using the J-space as part of its processing, not merely storing noise.
What it does not mean
The most obvious interpretation, that this is analogous to a human “inner voice” or “train of thought”, is precisely the one that interpretability researchers urge caution about.
“I don’t love using those kinds of terms,” Heaven told MIT Tech Review. “LLMs are not brains. Talking like this is misleading because it can suggest that LLMs are capable of more human-like things than they are.”
Anthropic itself has drawn an analogy to the space that some neuroscientists think our brains use to keep track of conscious thoughts. But the company’s official statement hedges carefully: “Drawing these analogies was helpful to us in designing our experiments, as they allowed us to make many non-obvious experimental predictions about the J-space that turned out to be true. At the same time, it’s important to note that there are some important differences between the J-space (and language models in general) and the human brain, so we don’t mean to claim there’s a perfect correspondence.”
The interpretability challenge
Understanding what goes on inside a large language model is exceptionally difficult. A medium-size model’s internal parameters, printed out, “would cover a city the size of San Francisco,” Heaven noted. Every output is the product of hundreds of billions of numerical operations, and the raw math, patterns of floating-point activations across millions of dimensions, looks like word salad.
Building the specialist tools needed to highlight specific activations at specific times requires prior understanding of the math. It is a circular problem: you need to know where to look before you can look.
What it could be useful for
The most practically promising application is monitoring. If J-space contains words that reveal a model’s internal state, including intentions that are not visible in the model’s final output, then monitoring J-space could catch undesirable behavior such as bias, sycophancy, or cheating before it manifests in generated text.
“This is one more step on the path to understanding this technology overall,” Heaven said, “rather than something that will be useful by itself.”
A narrative critique
Heaven also offered a critique of how the finding fits Anthropic’s broader brand. The company has positioned itself as the responsible AI developer, publicly warning about catastrophic risks and advocating for regulation. The J-space discovery, he noted, fits a convenient narrative: “They’ve built this really mysterious technology, but don’t worry, because they’re also the ones to figure it out.”
This narrative has real-world consequences. When Anthropic warned about model risks, the government responded by restricting development, a dynamic that benefited the company’s regulatory positioning. The J-space discovery, however scientifically legitimate, also reinforces the idea that Anthropic is uniquely positioned to understand and control its own creations.
For now, J-space is best understood as a genuine technical discovery whose practical significance remains uncertain. It adds a new tool to the interpretability toolkit, but it does not, yet, change what we know about how LLMs think. If they think at all.
Sources
O’Donnell J. “What Anthropic’s latest AI discovery does, and doesn’t, show.” MIT Technology Review (July 13, 2026). https://www.technologyreview.com/2026/07/13/1140343/what-anthropics-latest-ai-discovery-does-and-doesnt-show/