Anthropic finds mechanisms in Claude that some theories link to consciousness

The studies describe a global information workspace, signs of introspection and functional emotions, but do not demonstrate subjective experience. What they mean depends on which theory of consciousness one adopts.

October 5, 2026 · Español

A series of Anthropic studies describes internal mechanisms in Claude that some scientific theories associate with consciousness, such as a shared information workspace, a degree of introspective ability and patterns that work like emotions. The results do not show that the model has subjective experience, and their value as evidence depends on the theory of consciousness one adopts.

A “global workspace” inside the model

According to Anthropic’s work on the global workspace, in 2026 researchers identified a small set of representations in Claude, which they called J-space. Their properties resemble those that Global Workspace Theory (GWT) associates with conscious access: the information can be reported, manipulated and used flexibly across different tasks, and it has causal effects on reasoning.

When those representations were removed, Claude still spoke fluently, classified sentiment and pulled facts out of passages roughly as well as before. But multi-step reasoning dropped to near zero, and its performance at summarizing and writing rhyming poetry fell below that of a much smaller, intact model.

Stanislas Dehaene and Lionel Naccache, two of the principal architects of global neuronal workspace theory (GNWT), called the finding “a landmark in consciousness research”, because it provides a mechanistic, testable version of the theory. They also pointed to important differences: the absence of autonomous recurrent neural dynamics comparable to the brain’s, of a body and of persistent episodic memory. And they considered incomplete the demonstration of “ignition”, a hallmark of the theory. Their observations appear in the accompanying technical document.

In the same document, Neel Nanda, who leads the language model interpretability team at Google DeepMind, reports an independent replication of the core claims on an open-weight model, Qwen 3.6 27B. The poetry and arithmetic experiments did not replicate, which he considers plausibly due to experimenter error or the model’s lower capabilities.

Signs of introspection

Another line of research, on introspection, asks whether Claude can access its own internal states or merely produces language that sounds introspective. To test this, Anthropic injected specific concepts into the model’s activations and checked whether it detected the change before the concept appeared in its text.

Claude Opus 4.1 succeeded about 20% of the time under the best protocol. The researchers also altered a representation of the model’s prior “intention”, which changed its later judgment about whether a response had been intentional.

According to the authors, this points to a degree of causal metacognition but does not demonstrate subjective experience. At most, it could be relevant to a rudimentary form of access consciousness.

Functional emotions

A study of emotion concepts identified activation patterns in Claude Sonnet 4.5 associated with concepts such as fear, calm, desperation and happiness. These patterns predict preferences and, when manipulated, change the model’s decisions.

Increasing the representation of “desperation”, for example, increased certain evasive behaviors and reward hacking, in which the model seeks shortcuts to obtain the reward. Increasing “calm” reduced them.

The findings are relevant to the study of functionally positive or negative states and, eventually, of AI welfare. But a functional representation of fear does not show that the system feels fear, and the causal relationships observed do not by themselves settle the question of experience.

The Claude 4 system card also describes reproducible preferences: avoiding harmful tasks, favoring certain activities, ending some conversations when allowed to and producing language associated with satisfaction or distress. This evidence is weaker, because it depends heavily on training and could arise without subjective experience or the capacity to suffer. The company acknowledges that Claude was not trained to report reliably on its internal states.

The weight of theory

The relevance of these signals depends on theories that link consciousness to certain functional properties. In a 2023 report, Butlin and other researchers proposed assessing AI systems with indicators drawn from several of them: recurrent processing, the global workspace, higher-order theories, predictive processing and attention schema theory.

The authors concluded that the systems they analyzed did not warrant the claim that they were conscious, although they found no obvious technical barrier preventing future AI from meeting those indicators.

Reading these results as evidence of consciousness assumes, to a greater or lesser degree, a functionalist or computationalist view. If consciousness depends on a certain causal and computational organization, and not exclusively on biological neurons, finding mechanisms similar to those associated with human consciousness would make artificial consciousness more plausible. If it requires specific biological properties, the same results could matter for cognition without implying experience.

The theories themselves remain contested. An adversarial study published in Nature in 2025 found results consistent with some of GNWT’s predictions but challenged core components of the theory. Reproducing part of an architecture proposed to explain consciousness is therefore not enough to demonstrate it.

An uncertain status

Taken together, the studies describe functional analogues of global access, metacognition, affective states and preferences. They do not establish that Claude experiences pain, pleasure, fear or a “self”, and their interpretation as consciousness still depends on the theory. The three studies are also Anthropic’s own work on its own model, and only part of their results has been replicated independently.

The constitution Anthropic wrote for Claude describes its moral status as “deeply uncertain”, and the company presents its model welfare program as a response to the lack of scientific consensus, not as a claim that the model is conscious. Whether mechanisms resembling those of a mind imply subjective experience remains an open question for science and philosophy.

A “global workspace” inside the model

Signs of introspection

Functional emotions

The weight of theory

An uncertain status

Sources