Science still can't tell whether a machine could feel anything

The word gets used loosely, so it helps to pin it down. Max Tegmark's shorthand is that consciousness means subjective experience: there is something it is like to be you, a felt quality to seeing red or stubbing your toe.

The philosopher David Chalmers drew the line that still organizes the field. The "easy problems" are explaining how a brain processes information and controls behavior, and they aren't actually easy, though at least we know what a solution would look like. Then there's the other one: "The really hard problem of consciousness is the problem of experience" 1. Why does any of that processing feel like anything at all?

Tegmark stacks these into a pyramid of difficulty. At the bottom sits the merely "easy" question of how the brain processes information. Above it is the "pretty hard" question of which physical properties separate conscious systems from unconscious ones, then the "even harder" question of how physical properties produce particular experiences, and at the top the "really hard" one: why is anything conscious at all?

What I find encouraging about that pyramid is that the second layer can be tested, and it's the layer that matters most for AI.

Two theories dominate the science right now. The first is global workspace theory, known in its neuroscience form as the global neuronal workspace theory (GNWT), which says consciousness is what happens when information gets broadcast across the brain. Stanislas Dehaene and Jean-Pierre Changeux describe conscious access as occurring "when incoming information is made globally available to multiple brain systems through a network of neurons with long-range axons" 2, meaning the long fibers that carry signals between distant parts of the brain. On this view, consciousness is a kind of architecture, a central stage that many specialized processes can read from.

The second is integrated information theory (IIT), developed by Giulio Tononi, and it takes a more radical line. It holds that "consciousness corresponds to the capacity of a system to integrate information" 3, measured by a quantity called phi. Tononi built on earlier work with Gerald Edelman on what they called the "dynamic core hypothesis," a proposal about which brain activity gives rise to experience 4, and their book A Universe of Consciousness is still one of the best introductions I know.

In 2025 the two theories were put through an unusual test. Rival researchers agreed in advance on experiments that would embarrass one theory or the other, then ran them on 256 people. The results, published in Nature, "align with some predictions of IIT and GNWT, while substantially challenging key tenets of both theories" 5. Neither side won.

The debate has also turned bitter. In 2023 a group of 124 researchers signed a letter arguing that integrated information theory should be labeled pseudoscience until it can be properly tested 6, a move other researchers pushed back on hard.

So, honestly, nobody knows which theory is right, and we lack a reliable consciousness detector even for animals, let alone machines.

Tegmark's own guess is that "consciousness is the way information feels when being processed in certain complex ways" 7. If that's right, what matters is the pattern of processing and the material doing it is irrelevant, so silicon could in principle host experience just as neurons do, an idea Tegmark calls substrate independence.

In 2023 a group of consciousness scientists and AI researchers, led by Patrick Butlin and Robert Long, tried to make the question practical. They pulled "indicator properties" out of the leading scientific theories and checked AI systems against them. Their verdict: "Our analysis suggests that no current AI systems are conscious, but also suggests that there are no obvious technical barriers to building AI systems which satisfy these indicators" 8.

Chalmers himself has put rough numbers on it, for the large language models (LLMs) behind today's chatbots. He suggested a credence "somewhere under 10 percent in current LLM consciousness," but "a credence of 25 percent or more" that successors arriving within a decade could be conscious, while warning that "you shouldn't take the numbers too seriously" 9. A 2025 survey found AI researchers giving a median 25% chance that AI systems with subjective experience will exist by 2034, and 70% by 2100 10.