CAN ARTIFICIAL INTELLIGENCE EVER BECOME CONSCIOUS?

An AI can write about loneliness without establishing that it feels lonely. That gap is the heart of the consciousness debate. The question is not simply whether a machine can solve problems or describe an inner life, but whether there could be anything it is like to be that machine. Research offers ways to investigate the possibility. It does not yet provide a universally accepted machine-consciousness test, and neither a persuasive conversation nor a sweeping dismissal can substitute for one.

What would consciousness mean in an AI?

Here, consciousness means subjective experience: seeing a colour, feeling discomfort or having an experience with a point of view. Intelligence concerns capacities such as learning and reasoning. Self-description concerns what a system says about itself. These concepts can overlap without being identical. In his discussion of language models, philosopher David Chalmers examines possible architectural obstacles and future possibilities rather than treating verbal sophistication as a verdict. His argument is a philosophical assessment informed by science, not an experimental demonstration that any particular AI feels anything. David Chalmers: could a large language model be conscious?

Consider two different questions: can a system identify damage, and can it suffer? A sensor linked to a warning message answers the first. It does not, by itself, answer the second. Conversely, the absence of a human face would not logically rule out experience. The approved hero makes the encounter easy to imagine, but its humanoid body is a visual metaphor. Shape, voice and apparent vulnerability can influence an observer without establishing the internal processes that competing theories consider relevant.

From impressive behaviour to testable indicators

A 2023 report by Patrick Butlin and colleagues proposed looking for computational properties derived from scientific theories of consciousness. Their candidates included recurrent processing, the sharing of information through a global workspace and higher-order representations of a system’s own states. The report judged the systems it examined unlikely to be conscious. That conclusion belongs to its assessed systems and theoretical assumptions; it is not a timeless survey of every model released afterward. Importantly, satisfying a proposed indicator is evidence within an approach, not a certificate of experience. Butlin and colleagues: theory-derived indicators, 2023

Their 2025 peer-reviewed framework develops this theory-based strategy. Its value is methodological: ask what is happening inside a system, specify why that property should matter, and consider several explanations. A component labelled memory or attention is not automatically equivalent to the mechanism a consciousness theory describes. Researchers must investigate its actual operation and causal role. This is more demanding than counting capabilities, because impressive performance and the proposed basis of experience are different targets of explanation. Butlin and colleagues: identifying indicators of consciousness, 2025

The theories do not all make the same claim

Global-workspace approaches emphasize making selected information broadly available to other processes. Recurrent-processing approaches emphasize feedback interactions. Higher-order approaches emphasize representations of mental states. These are research programmes with distinct commitments, not interchangeable names for a sufficiently large neural network. Computational functionalism adds a philosophical assumption: the right organization of processes could matter more than the material implementing them. Someone who thinks living biological processes are indispensable will assess a digital imitation differently. Naming the assumption makes the disagreement clearer; it does not resolve it. Butlin and colleagues: theory-derived indicators, 2023

What a major human experiment actually tested

The Cogitate Consortium’s 2025 Nature study tested divergent predictions of global neuronal workspace theory and integrated information theory in 256 human participants. It used several recording methods while people viewed visible stimuli. Results supported some predictions and challenged important claims of both theories. This was progress in testing specified accounts of human experience, not a contest won by an AI or proof that consciousness is beyond science. Transferring an indicator from human brains to a different kind of system requires further argument and validation. Cogitate Consortium: adversarial theory testing, 2025

A newer model still expresses uncertainty

The Digital Consciousness Model report, first posted in 2026, combines multiple theoretical perspectives in a probabilistic framework. Its initial assessment found evidence against consciousness in the 2024 language models it considered, while describing that evidence as inconclusive. This is a preprint-based assessment, not an agreed diagnostic standard. Its dates matter: a report updated in 2026 does not automatically evaluate every 2026 system. Probabilities also depend on choices about theories and evidence, so they should not be presented as direct measurements of a machine’s private experience. Shiller and colleagues: Digital Consciousness Model, 2026 preprint

What would stronger evidence look like?

A serious investigation would disclose which architecture is being assessed, define indicators before interpreting results and test alternative explanations. It would examine internal mechanisms alongside behaviour, rather than accept a system’s declarations as decisive. Changing a relevant mechanism and observing the consequences could strengthen a causal explanation, although interpretation would still depend on the theory. Agreement across independent methods would be more informative than one dramatic transcript. This is a research agenda, not a promise that a simple checklist will finally settle subjective experience. Butlin and colleagues: identifying indicators of consciousness, 2025

The difference between a prerequisite and a guarantee is especially important. If a theory says a process is necessary, finding that process does not establish that it is sufficient. Nor does an unexplained capability automatically favour consciousness over a non-conscious account. A useful claim should identify what observation would weaken it. Otherwise, any surprising response becomes confirmation, and the proposed test stops distinguishing between explanations. Intellectual openness works best when paired with the possibility of being wrong.

Ethics without pretending the debate is settled

Two errors deserve attention: attributing experience where there is none, and overlooking experience if it becomes possible. The first can mislead users and shift responsibility away from the people deploying a system. The second could matter for future welfare. Caution does not require treating every chatbot as a person. It requires honest descriptions of evidence and uncertainty. Whether an AI can experience something and whether it can control an action are also different questions, explored in our article on free will and decision-making.

The most defensible answer is conditional. Some philosophical and scientific approaches allow artificial consciousness in principle; others challenge what a digital implementation could achieve. The cited research does not establish conscious AI, and it does not demonstrate an impossibility for all future machines. Progress will come from sharper theories and better tests, not from mistaking a familiar face for an inner life. For the wider problem of what counts as an ultimate explanation, read why anything exists.

Related reading: What brain experiments can tell us about free will.