Can AI Be Conscious? What the Science Says Now

Can AI be conscious? For the AI systems that exist today, the answer from a major 2026 assessment is no. In 2026, 20 philosophers, neuroscientists and AI researchers published Identifying indicators of consciousness in AI systems in Trends in Cognitive Sciences, and one of them, Colin Klein, summed up its view in June: no existing AI system, ChatGPT included, is conscious, but nothing bars AI systems from becoming conscious in future.
In September 2026 a second group, including Anil Seth, Murray Shanahan and Shane Legg, showed how much that verdict rests on theory. In From cacophony to hierarchy, their illustrative estimates for current language models ran from below 1% to about 80%, depending on which theory of consciousness is trusted. The authors stress that this is not a measured probability. It shows that the disagreement is about theory more than data.
How researchers ask whether AI can be conscious
There is no agreed test that detects consciousness from the outside, not even in animals. Researchers therefore work from theories: each major theory says which kind of processing matters, and that can be checked against how an AI system is built.
The method was set out in a 2023 report, Consciousness in Artificial Intelligence, which turned leading theories into a list of "indicator properties" and concluded that no AI system of that time was conscious, while finding no obvious technical barrier to building one that has the indicators. The 2026 paper, whose authors include Patrick Butlin, Robert Long, Yoshua Bengio, Jonathan Birch and David Chalmers, develops that method. It treats the indicators as evidence that raises or lowers how likely a particular system is to be conscious, rather than as a pass or fail test, and Klein notes that the useful indicators are all structural: they concern how a system is built, not how it behaves. The guide on how to test if AI is conscious walks through the method.

Why today's chatbots are judged unlikely to be conscious
In 2023 David Chalmers argued in Could a Large Language Model be Conscious? that language models probably are not conscious, because mainstream theories point to features they lacked: recurrent processing, a global workspace and unified agency. He added that these gaps could plausibly close within about a decade.
Since then, researchers have started to look for those features inside the models themselves:
- A workspace inside Claude. In July 2026 Anthropic researchers reported a "J-space" in Claude models: a small set of representations the model is poised to put into words, which they found has many of the functional properties of the brain's global workspace. They take no position on whether it involves experience. The theory behind it is explained in global workspace theory and AI.
- Limited self-monitoring. In October 2025 Anthropic's introspection experiments found that Claude Opus 4.1 noticed concepts planted in its own activity only about 20% of the time, and Anthropic said this does not tell us whether Claude is conscious.
Both findings concern what a model does with information. Neither shows that there is anything it is like to be the model.

Why the question is still open
First, the theories disagree about what consciousness needs. Some are stated in terms of function, so a machine that does the right things would count. Others tie consciousness to physical structure. Researchers working on Integrated Information Theory argued in 2024, in Dissociating Artificial Intelligence from Artificial Consciousness, that a computer could match a conscious system's behaviour exactly without sharing its experience. The September 2026 framework makes the same point with numbers: the verdict on current language models depends as much on where you place your trust among theories as on how you read the evidence. Our guide to integrated information theory and AI explains the structural view.
Second, AI developers now treat the question as open in their own rules. In January 2026 Anthropic published a new constitution for Claude that expresses uncertainty about whether Claude might have some kind of consciousness or moral status, and says the company cares about Claude's wellbeing.
A closer look: "not now" is not "never"
The 2026 verdict is a statement about today's systems, judged against theories that may themselves be wrong. The September 2026 framework adds a reason to keep watching: the features its indicators look for overlap substantially with the features needed for general intelligence, so more capable AI may become a stronger candidate. Its authors recommend what they call structured agnosticism: state the theory you are relying on, update as evidence comes in, and give probabilities rather than verdicts.
That is why researchers who doubt today's systems still argue for preparing now, as the guide to AI model welfare describes. The words matter too: "conscious" and "sentient" are not the same, as AI sentience vs consciousness explains.
What to keep in mind when you read a claim about conscious AI
- How recent is the research? The 2023 assessments have been revised; check the date of the work behind a claim.
- Which theory is it using? A verdict only makes sense relative to some account of what consciousness is.
- Is it about today's systems or future ones? Most careful verdicts are confident about the first and cautious about the second.
- Does it rest on what the AI says? A fluent answer about feelings is not evidence of feelings.
- Is it a probability or a verdict? Current research increasingly speaks in likelihoods.
Related reads
- How to test if AI is conscious: indicators and probabilities in detail.
- Global workspace theory and AI: the theory behind the workspace found in Claude.
- AI sentience vs consciousness: two words that are often confused.
- The moral status of AI: why the answer matters ethically.
More in The Big Question · Report a correction