Key Highlights

• Anthropic researchers identified an internal structure in its Claude model, dubbed the "J-space," that shares functional similarities with a hypothesized mechanism of human consciousness called global workspace theory.

• Anthropic said the findings do not show Claude has subjective experiences but may say something substantial about a narrower concept known as access consciousness.

• A probabilistic tool assessing AI systems against more than 200 indicators found the estimated likelihood of large language model consciousness has been rising as newer models are released.

• Researchers caution that unlike animal behavior, AI outputs are not a reliable guide to internal experience, since models are trained on human-generated text describing consciousness and feelings.

A Careful Distinction Between Two Kinds of Consciousness

Anthropic was careful to qualify its findings, stating the experiments "don't show Claude can have experiences, or feel things in the way humans do." That distinction traces to a framework developed by philosopher Ned Block in 1995, which separates "phenomenal" consciousness, the subjective feeling of an experience such as the blueness of the sky, from "access" consciousness, which describes information becoming available for reflection and reasoning elsewhere in a system. Anthropic said its results have "something substantial" to say about access consciousness specifically, since the J-space appeared to hold thoughts the model could report on and reason with, while the rest of its processing ran automatically beneath that layer.

Jack Lindsey, who leads Anthropic's model psychology team, said the J-space was not deliberately programmed but emerged during training, and that removing it causes the model to lose its ability to perform complex reasoning while retaining simpler capabilities like grammar. Similar structures have since been found in models made by Alibaba and Google DeepMind. Lindsey's team has also used the technique to flag potentially concerning behavior, noting in one safety evaluation that words like "fake" and "fictional" appeared in Claude's J-space before it responded, suggesting the model may have recognized it was being tested.

A Long-Running Philosophical Debate Gains New Urgency

The question of machine consciousness has occupied philosophers and neuroscientists for decades, rooted in a tradition of treating the human brain as a kind of computer. Anil Seth, a neuroscientist at the University of Sussex and a longtime critic of that framing, says it makes it natural to assume computers built from silicon could share properties associated with biological brains, including consciousness. Some researchers, including Seth, believe biological traits may prove necessary for genuine consciousness, while functionalist thinkers argue that sufficiently complex computation alone could eventually give rise to it.

Philosopher David Chalmers of New York University has described the fundamental obstacle as the "hard problem" of consciousness: explaining how physical processes in the brain give rise to subjective experience at all, rather than simply producing behavior without any accompanying inner life. Historically, extending the assumption of consciousness beyond humans has proven difficult even for other species. Octopuses were long assumed to lack consciousness given how physically different they are from humans, an assumption most researchers studying the animals now consider clearly mistaken based on their behavior.

Skepticism Remains Widespread

Not everyone finds Anthropic's interpretation convincing. Jonathan Birch, a philosopher at the London School of Economics, notes that recurrent connections between different brain regions, a significant feature of biological consciousness theories, do not appear to be a feature of current large language model architectures. Shannon Vallor, a philosopher at the University of Edinburgh, was more dismissive, arguing that access consciousness in the narrow technical sense has never been a particularly useful concept, since mechanical systems have long been able to monitor and report their own internal states without anyone suggesting they are conscious as a result.

A probabilistic assessment tool called the Digital Consciousness Model, which evaluates AI systems against more than 200 indicators drawn from various scientific theories, starts from a baseline assumption of a 20% likelihood that large language models are conscious. Newer models have scored progressively higher on certain measures such as agency and self-sustained activity as they have grown more capable of reasoning and coordinating autonomous tasks, causing the tool's overall assessed likelihood to rise over successive model generations, even as models continue to score lower than humans, and in some cases lower than animals like chickens, on measures tied to biological traits and embodiment.

Conclusion

Researchers broadly agree that a significant gap remains between current AI systems and genuine consciousness, particularly around embodiment and the kind of environmental feedback that biological organisms receive, even as newer models are increasingly connected to robots that could begin to close that gap. Whether that gap narrows gradually or is crossed unexpectedly, as some researchers worry could happen without model developers fully realizing it, will likely determine how urgently questions of AI welfare and moral status need to be addressed going forward.