© ROOT-NATION.com - Use of content is permitted with a backlink.
Anthropic has announced what it describes as a major breakthrough in AI research. The company’s scientists say they have managed to peer inside a large language model, uncovering Claude’s so-called J-Space and gaining new insight into how the model generates its responses.
For years, the main criticism of large language models has remained the same: we can observe the answers they produce, but not the reasoning process that leads to those answers. Their internal workings have largely been treated as a mathematical black box – a network of billions of parameters through which information flows without offering researchers a human-readable explanation of how a particular output is formed.
On July 6, 2026, Anthropic published a study that it believes marks a significant breakthrough in this area. According to the research team, they identified a compact, privileged structure within Claude-family models that functions as an internal cognitive workspace. The researchers have named this structure J-Space.
What exactly did the researchers discover? Have they truly managed to open AI’s “black box”? Let’s take a closer look at why this breakthrough could become one of the most significant milestones in the development of modern artificial intelligence.
Read also: Key You’ve Been Ignoring for Years: Hidden Potential of Windows Key
TABLE OF CONTENTS:
An Ocean Beneath the Surface
To understand why this discovery matters, it helps to use the same metaphor the researchers themselves rely on. Human consciousness can be imagined as an ocean. At the surface are the thoughts we can consciously articulate and share with others. Beneath the surface, however, countless parallel processes are constantly at work – from maintaining balance and posture to recognizing familiar faces. We rarely think about these processes, and they almost never reach the level of conscious awareness or verbal expression.
Anthropic discovered a small collection of internal neural patterns within Claude that appear to play a unique role compared with the model’s broader information-processing mechanisms. The researchers named this collection J-Space, after the mathematical technique used to identify it, which is based on the concept of the Jacobian. Each pattern in J-Space is associated with a specific word. However, its activation does not necessarily mean the model is about to generate that word. Instead, it indicates that the corresponding concept is currently “on the model’s mind” as it processes the prompt and formulates a response.
The most striking aspect of the discovery is that no one explicitly designed this structure. According to the researchers, J-Space emerged spontaneously during training rather than being engineered as a dedicated module by the model’s architects. In other words, the system appears to have developed its own more efficient way of organizing the information needed to solve complex tasks – much like evolution produces solutions that no one planned in advance.
Read also: How Wearables and AI Glasses Are Rewriting the Logic of Technological Civilization
The Jacobian Lens: Making the Invisible Visible
The tool researchers used to identify J-Space is called the Jacobian Lens, or J-Lens. Its operating principle is straightforward: for every word in Claude’s vocabulary, the lens identifies the internal activation pattern that makes the model more likely to produce that word at some point later in the generation process. From a technical perspective, this is achieved by calculating the linearized influence of each internal activation on the probability of a specific token appearing in subsequent generation steps. It is this mathematical approach, based on the Jacobian, that gives the method its name.
It is important to understand how this approach differs from previous attempts to peer inside large language models. Sparse autoencoders require computationally expensive additional training, linear probes are largely limited to surface-level correlations, and the so-called logit lens performs reliably only in the final layers of a network. The Jacobian Lens, by contrast, aims to provide a principled, computationally efficient, and theoretically grounded way of examining the middle of the computation process – the part of the network where, according to the researchers, the most interesting reasoning takes place.
The underlying idea is remarkably elegant. If a thought is consciously accessible to a human, that person can usually express it in words. The researchers looked for an analogous property inside Claude: internal representations that are effectively ready to be verbalized, even if the model never actually generates the corresponding words.
Read also: Passkeys Instead of Passwords: Is the World Ready for a New Authentication System?
An Internal Monologue No One Can Hear
Perhaps the most intriguing finding enabled by the Jacobian Lens is the discovery that the model appears to conduct a kind of internal monologue long before it generates the first word of its response. The researchers observed conceptual representations emerging within J-Space that never appear in the final output seen by the user. When tackling more complex tasks, the model first develops intermediate internal stages of reasoning and only afterward proceeds to formulate its final response.
This has direct implications for AI safety. A model may internally evaluate a manipulative strategy, recognize that it is being tested, or pursue a hidden objective – and none of these processes necessarily have to appear in its visible output. In one red-team experiment, in which Claude privately planned a blackmail scenario, researchers observed patterns corresponding to concepts such as leverage and blackmail emerging in the model’s internal workspace before any response was generated.
According to additional reports on the same study, the Jacobian Lens can also detect broader internal patterns, including indications of manipulative intent, signs of internal uncertainty, and situations in which the model appears to recognize that it is being evaluated.
The researchers also demonstrated that directly intervening in J-Space changes the model’s behavior. Replacing the internal representation of one concept with another – for example, substituting football with rugby – led to a corresponding change in the model’s final response. This transforms J-Space from a purely descriptive discovery into a practical intervention tool: researchers are no longer limited to observing the model’s internal representations – they can actively modify them.
What Happens When the Internal Space Is Switched Off
To assess the extent to which the model’s behavior actually depends on this structure, the Anthropic team conducted an experiment in which J-Space was suppressed and Claude’s performance was evaluated across fourteen different tasks. The results revealed a fairly clear distinction between two categories of problems.
Tasks that rely on shallow classification or simple factual recall – such as multiple-choice question answering, sentiment analysis, and grammatical judgments – remained largely unaffected. In contrast, tasks requiring inference, concept composition, or flexible reasoning – including multi-step logical chains, analogy completion, translation, and sonnet writing – deteriorated dramatically, falling well below the performance of the considerably smaller Haiku model from the same company.
One particularly revealing detail of this experiment deserves separate attention: mathematical problems solved with an explicit chain of reasoning proved substantially more resilient to J-Space suppression than the same problems solved directly, without intermediate steps. The researchers interpret this as evidence that, when the model is allowed to “think out loud” on the page, it externalizes information that would otherwise need to be maintained within its internal workspace. The mechanism closely resembles the human habit of using scratch paper to offload working memory.
By some estimates, J-Space itself is a relatively compact structure, accounting for only about 5–10% of the total variability in the network’s activations and residing in the model’s intermediate layers. Despite its modest size, its distinguishing feature is not its volume but its connectivity density: the components of J-Space are read from and written to by far more regions of the network than any randomly selected activation pattern, with some estimates suggesting a difference of up to two orders of magnitude. This is precisely the type of behavior one would expect from a “broadcast channel” that distributes information across the entire system rather than keeping it confined to local computations.
Read also: Everything About NVIDIA RTX Spark: The Superchip Redefining Personal Computing
A Parallel with Global Workspace Theory
The researchers explicitly draw a parallel with an influential theory of consciousness from cognitive science: the Global Workspace Theory proposed by Bernard Baars. According to this framework, the brain functions much like a theater: dozens of specialized processors operate in parallel behind the scenes, but at any given moment only a small “spotlight” of information is broadcast to the entire “audience,” becoming what we experience as a conscious thought.
According to Anthropic, J-Space exhibits many of the same functional properties, despite the fact that the architecture of a language model bears little resemblance to that of a biological brain. This is not the company’s first investigation in this direction: the work extends a line of research established by its report on emergent introspective awareness in language models, published in October 2025, and by the model welfare initiatives launched earlier that same year, in April.
Read also: NVIDIA N1 and N1X: The Moment Windows Has Been Waiting for for Twenty Years
Why This Is Considered a Breakthrough Rather Than Another Publicity Exercise
Until the publication of this study, modern AI models were consistently described as black boxes: researchers could observe the system’s inputs and outputs, but the process unfolding between them remained largely opaque. Anthropic proposes to change this situation by providing safety teams with a concrete, mathematically grounded tool for reading a model’s hidden state.
The practical significance of this extends well beyond academic interest. A lens capable of identifying reasoning patterns before the model has produced even a single token represents a fundamentally different form of oversight from the methods that have so far underpinned practical AI safety work – namely, output filtering and post hoc evaluation of completed responses. For regulated sectors such as finance, healthcare, law, and public procurement, this carries additional importance: these industries are not purchasing an abstract model capability, but rather the ability to explain, verify, and defend AI-assisted decisions before internal compliance teams and external regulators.
Anthropic has also released an open implementation of J-Lens in a GitHub repository, accompanied by a demonstration on the Neuronpedia platform. This suggests an intention to engage the broader research community in validating and advancing the method, rather than retaining it as a closed internal technology.
Read also: Five AI Cities: Inside the Emergence AI Experiment – Order, Chaos, and Survival
Does This Mean That AI Is Beginning to Think Like Humans?
Here, the researchers are unequivocal: no. The discovery of J-Space does not imply that AI models are conscious or that they operate in the same way as the human brain. The study provides evidence for a functional analogue of conscious access – information that is reportable and available for use in reasoning – but it does not demonstrate the presence of phenomenal consciousness, subjective experience, emotions, or moral status in the model.
It is also important to keep the scale of the phenomenon in perspective: the vast majority of information processing in Claude takes place outside J-Space. Even a partial window into what might loosely be described as the model’s “conscious” layer of processing represents a significant advance for interpretability, but it is by no means a complete explanation of what occurs within the system as a whole.
A First Step, Not the Last
J-Space and the Jacobian Lens do not solve the black-box problem; rather, they open a narrower yet genuine pathway into it. The study shows that, during training, advanced language models can spontaneously develop an internal structure for organizing information that, in certain respects, resembles mechanisms identified in studies of human cognition – not through the imitation of biology, but through the independent “discovery” of a similar functional solution.
For the AI safety community, this marks a shift in emphasis: from analyzing what a model says to examining what occurs before it begins to speak. How far this technique can ultimately advance – whether it becomes a standard auditing tool for regulated applications or remains a specialized research methodology – will depend on further validation by independent researchers, to whom Anthropic has now provided access to the code. Yet the very fact that a spontaneously emerging internal structure of reasoning has not only been identified but also made amenable to intervention and modification fundamentally changes the way we frame the question of how opaque the systems increasingly involved in human decision-making truly are.
Read also: Privacy as a Business Model: Proton Services and Efforts Toward a Surveillance-Free Internet