The remarkable behaviour of language models makes machine consciousness a legitimate scientific question. It does not justify confusing linguistic competence, internal computation or self-reference with subjective experience.
Large language models can discuss fear, pain, love, uncertainty and even their own supposed internal states. They can write “I think”, “I feel” or “I am afraid” with extraordinary fluency. As these systems become more capable, an obvious question follows: could some of them already be conscious?
The scientifically defensible answer today is neither a confident “yes” nor a metaphysical “never”. It is more precise: there is currently no scientific demonstration that today’s LLMs possess subjective experience, and science itself does not yet have a single broadly accepted and decisively validated theory explaining how human consciousness arises.
Consciousness science has not converged
The first difficulty appears before artificial intelligence even enters the picture. Neuroscience has made impressive progress in identifying neural processes associated with conscious perception. But identifying what happens in the brain when experience occurs is not the same as explaining why particular physical processes are accompanied by experience at all.
Several major theoretical approaches remain active. Global neuronal workspace theories associate conscious access with information becoming globally available across specialised neural systems. Integrated information theory focuses on a system’s intrinsic causal and informational integration. Higher-order theories emphasise representations of mental states. Other approaches stress recurrent processing, predictive mechanisms, embodiment, interoception or the biological organisation of living organisms.
These are not merely different vocabularies for the same explanation. They disagree about which mechanisms matter and about what properties are necessary or sufficient for consciousness.
A major adversarial collaboration published in Nature in 2025 illustrates the problem. Researchers jointly designed preregistered experiments to test divergent predictions of two leading theories, integrated information theory and global neuronal workspace theory. Some predictions were supported, but important claims of both theories were also challenged.
That is exactly how science should advance. But it also means we cannot truthfully say that consciousness science has already established a theory that can simply be applied to a transformer and return a reliable answer: conscious or not conscious.
Access is not the same thing as experience
Much of the confusion surrounding AI consciousness comes from using the word “consciousness” for different phenomena.
Access consciousness is functional. Information is available for reasoning, reporting, planning and the control of behaviour. Phenomenal consciousness refers to subjective experience itself. Pain does not merely provide information about bodily damage. It hurts. Seeing red is not merely successful classification of a wavelength. There is something it is like to see red.
The distinction matters enormously for LLMs.
A computational system might have internal information that can be retrieved, manipulated, reported and used to guide later computation. It could even contain a model of its own processing. None of this, by itself, establishes that there is something it feels like to be that system.
What Anthropic actually discovered
Anthropic’s July 2026 research is important precisely because it moves beyond superficial chatbot behaviour and examines internal mechanisms.
Using a technique called the Jacobian Lens, researchers identified a relatively small, privileged set of internal representations in Claude, which they call the J-space. These representations appear to play a special computational role. They can carry intermediate information used in reasoning, influence later processing and represent concepts that the model has not expressed in its visible output.
When researchers intervene on these representations, some complex reasoning abilities are strongly disrupted while basic linguistic abilities can remain comparatively intact. This suggests that the J-space is not simply a mirror of the next word Claude is about to produce. It functions as an internal workspace used by multiple computations.
One particularly revealing experiment involved a question asked in Chinese. Intermediate English representations such as “big” and “bigger” appeared while the model was preparing a Chinese answer. Changing those English representations causally changed the eventual Chinese output. This matters for multilingual interpretability as well as AI research more generally, because the internal representational structure of a model may not be linguistically neutral.
These are significant findings.
They are not evidence that Claude feels anything.
Anthropic explicitly states that its experiments do not show that Claude has experiences or feels in the human sense. The researchers instead argue that the J-space has functional similarities to mechanisms associated with conscious access.
Functional similarity does not establish subjective experience
This is closely related to a broader error in discussions of artificial intelligence: assuming that reproducing a function means reproducing the nature of the entity that performs it.
An aircraft reproduces the function of flight without reproducing a bird. A chess engine can outperform every human chess player without feeling ambition, frustration or satisfaction. An LLM can write poetry, reason through a legal problem and generate detailed descriptions of grief without that behaviour demonstrating grief, a continuous self or lived experience.
None of this proves that machine consciousness is impossible.
Some researchers argue that the right functional or computational organisation might be sufficient. Others challenge that assumption. Anil Seth, for example, argues from biological naturalism that consciousness may depend deeply on properties of living, self-maintaining and embodied organisms. On this view, current AI development is not necessarily travelling along a path that leads to consciousness merely by increasing model size and capability.
The important point is not that one side has already won. It is that the underlying scientific disagreement remains unresolved.
Why asking an LLM whether it is conscious does not solve the problem
LLMs are trained to generate contextually appropriate language. They have absorbed enormous quantities of human writing about minds, emotions and consciousness. They can therefore produce persuasive first-person descriptions.
A statement such as “I am experiencing fear” is an output. It becomes evidence of subjective experience only if we already possess a justified theory connecting that kind of output and the underlying mechanism to consciousness.
But that is precisely what we do not yet have.
Self-report is useful evidence in humans because it sits within a much larger body of knowledge about shared biology, nervous systems, behaviour, development and bodily life. Extending the same inference directly to an artificial language model would simply assume the conclusion we are trying to establish.
Open models are part of scientific verification
There is another important consequence of the Anthropic work.
Claims about hidden properties of AI systems are scientifically useful only if they can be challenged and replicated. Anthropic released an open implementation of the Jacobian Lens and made the method applicable to open-weight models. This allows researchers outside a single company to test related hypotheses across architectures, training regimes and languages.
This is why openness matters for AI science.
Open-weight models, open-source interpretability tools, public evaluation datasets and reproducible experiments are not merely licensing preferences. They are infrastructure for independent verification.
For Greek AI research in particular, this creates a concrete opportunity. Rather than debating machine consciousness through screenshots of chatbot conversations, researchers can test interpretability methods on open models using Greek prompts, Greek corpora and multilingual evaluation sets. They can ask whether internal structures identified in English behave similarly in Greek and whether interpretability techniques themselves contain linguistic biases.
Neither mythology nor denial
Artificial intelligence is not a natural force advancing independently of human decisions. We design these systems, select their training procedures, build the infrastructure they run on and decide where they are deployed.
The language we use to describe them matters.
If we prematurely call LLMs conscious, we risk transforming interesting computational properties into anthropomorphic mythology. If we dogmatically declare machine consciousness impossible under all conceivable conditions, we close a scientific question before possessing the knowledge required to settle it.
A more demanding position is also a more scientific one: take the evidence seriously, but do not claim more than the evidence demonstrates.
We now know that LLMs can develop remarkably sophisticated internal computational structures. We do not know that they experience the world.
Until we understand both human consciousness and artificial systems far better, that distinction is not semantic caution. It is a basic requirement of scientific integrity.