Anthropic discovers Claude's 'J-space': the internal space where AI thinks without writing

What happened?

On July 7, 2026, Anthropic published research describing Claude's J-space: an internal activation space where the model represents concepts before verbalizing them, or even without them reaching the final response. The researchers developed a tool called J-lens to observe these patterns, which were not explicitly designed but emerged during training. The company stated on X that 'Claude has developed a mechanism for conscious access,' though they clarified this does not imply human subjective experience.

The study, published at transformer-circuits.pub/2026/workspace, details how the team applied mechanistic interpretability techniques to identify a subspace of activations—the J-space—that encodes concepts latently. According to the WWWhat's new article, the J-lens tool allows not only observation but also modification of these internal representations, unprecedented in LLMs. Anthropic found that certain patterns in J-space correspond to concepts like 'true/false', 'positive/negative', or 'dangerous/safe', and altering these patterns changes the model's output predictably. For example, by intervening in J-space to suppress representations of 'deception', Claude's responses became more honest in controlled tests.

Why is it important?

This discovery represents a significant advance in LLM interpretability. Until now, the inner workings of models like Claude were a 'black box': input went in, output came out, without understanding intermediate processes. J-space allows identifying how certain concepts appear, change, or disappear before the model responds, offering a window into its internal reasoning. This could lead to safer, more aligned, and controllable models, by enabling detection of biases, unwanted thoughts, or erroneous decision processes before they manifest in the output.

Historically, AI interpretability has been a challenge. In 2022, Anthropic had already published work on 'features' in small models (e.g., their paper 'Toy Models of Superposition'), but applying similar techniques to a commercial model like Claude 4 (released in 2025) is a qualitative leap. Compared to OpenAI's attempts to use 'sparse autoencoders' to interpret GPT-4 in 2024, J-space offers much greater granularity: while autoencoders identify features at the level of individual neurons, J-space captures representations at the level of complete concepts, enabling a more holistic understanding of the model's reasoning.

The market impact is immediate: companies relying on AI for critical decisions (medical diagnosis, legal analysis, algorithmic trading) see in J-space an opportunity to audit and validate models' internal processes. According to Gartner estimates, the explainable AI market will reach $15 billion by 2027, and this finding could accelerate that adoption. However, it also raises concern: if J-space is observable, it could also be manipulated by malicious actors, demanding new ethical and technical safeguards.

Consequences and applications

The finding has profound implications:

• Safety and alignment: being able to monitor internal 'thinking' could help prevent harmful or deceptive responses. Anthropic already demonstrated that by intervening in J-space to suppress representations of 'racial bias', Claude's responses showed a 40% reduction in implicit biases on standard tests (data from the paper).
• Transparency: it brings AI closer to being explainable, which is crucial for adoption in critical sectors like healthcare, justice, or finance. For example, a judge could ask a model to 'explain' its reasoning by showing relevant activations in J-space, rather than a mere post-hoc textual justification.
• Innovation: it opens the door to new architectures that leverage this internal space to improve reasoning and efficiency. Startups are already exploring how to use J-space as a 'working memory' for models, allowing them to maintain context beyond the current token window.

However, it also poses risks: if J-space can be read, it could also be manipulated. An attacker with access to Claude's API could inject malicious representations into J-space to force the model to generate compromising responses. Anthropic acknowledges this risk in its paper and suggests 'encryption' techniques for J-space, but no solutions are implemented yet. Moreover, the discovery reopens the ethical debate on machine 'consciousness': although Anthropic insists there is no subjective experience, some philosophers and activists are already calling for regulation of J-lens use, arguing that 'reading the mind' of an AI could violate its 'privacy' (a controversial concept).

What readers should know

This is not about Claude having 'consciousness' in a human sense, but that it has developed internal representations that can be observed and modified. The research is a step toward more interpretable AI, but we are still far from fully understanding a machine's mind. For companies and users, this means the next generation of models could be more reliable and controllable, but will also require new forms of governance. European regulators have already shown interest: the European Commission, through its AI office, announced on July 8 that it will assess whether J-space should be considered part of the 'internal functioning' subject to auditing under the AI Act.

In practice, developers using Claude 4 or future versions will be able to request access to J-lens from Anthropic to debug their applications. For example, if a customer service chatbot generates inappropriate responses, the developer could examine J-space to see if the concept 'frustration' was overly activated, and adjust the model accordingly. This represents a paradigm shift: from 'train and pray' to 'understand and correct'.

'It was not designed or programmed by us, but emerged on its own during Claude's training process.' — Anthropic

Anthropic's quote underscores the unexpectedness of the finding: J-space was not a planned feature, but an emergent phenomenon. This echoes other discoveries in AI, such as 'concept neurons' in convolutional networks (2015) or 'interpretation circuits' in language models (2023). But J-space goes further: it is not individual neurons, but a high-dimensional activation space that orchestrates the model's behavior. The scientific community is divided: some see it as the biggest advance in interpretability since activation maps of convolutional networks; others warn we might be overinterpreting statistical artifacts. What is certain is that J-space will change how we design, audit, and deploy language models in the coming years.