In 1998 Christof Koch bet David Chalmers that science would find the neural correlate of consciousness within 25 years. In 2023 he paid up in Madeira. Meanwhile the two leading research programs reach opposite verdicts on the same machine, so any claim about AI consciousness has to answer one question before it means anything: on which theory?
Christof Koch carried a wooden box full of wine onto the stage and pulled out a 1978 bottle of Madeira for David Chalmers. It was a muggy June night in 2023 in Greenwich Village, and more than 800 neuroscientists, philosophers and curious members of the public had packed the auditorium, The New York Times reported. Koch was paying off a bet.
The two men had made it in June 1998, late one night in a bar in Bremen, Germany, where a conference had brought them. Koch, a neuroscientist, had worked for years with Francis Crick on a quest for what they called the "neural correlate of consciousness," and he bet that scientists would find one within 25 years. Chalmers, a philosopher, took the bet. The prize was a few bottles of fine wine. "When you're young, you've got to believe things will be simple," Koch told the audience.
Then he offered a new bet, double or nothing: a brain marker of consciousness by 2048. Chalmers shook on it at once. "I hope I lose," he said. "But I suspect I'll win."
Chalmers had reason to be confident. In 1995 he split the problem of consciousness in two, and he argued that the half he called hard is one no explanation of function can reach. Most arguments about machine consciousness turn on that split, so it is worth understanding in detail.
Chalmers set out the split in "Facing Up to the Problem of Consciousness," in the Journal of Consciousness Studies in 1995. The easy problems are those of explaining functions: "the ability to discriminate, categorize, and react to environmental stimuli; the integration of information by a cognitive system; the reportability of mental states," along with a system's access to its own internal states, the focus of attention and the deliberate control of behavior. Easy is relative. Getting the details right, he guessed, "will probably take a century or two of difficult empirical work." But they are problems of a familiar kind. To explain a function, he wrote, "we need only specify a mechanism that can perform the function."
Every item on that list is something an engineer could write a test for. A system either discriminates the stimuli or it doesn't, and it either reports its internal state accurately or it doesn't. The hard problem is the item with no test. "The really hard problem of consciousness is the problem of experience," Chalmers wrote. "When we think and perceive, there is a whir of information-processing, but there is also a subjective aspect." Borrowing the phrase from Nagel's 1974 paper "What Is It Like to Be a Bat?", he described it as there being "something it is like to be a conscious organism": the felt quality of redness, "the sound of a clarinet, the smell of mothballs." Then he asked the question behind this essay's title: "Why should physical processing give rise to a rich inner life at all? It seems objectively unreasonable that it should, and yet it does."
The claim is precise. You could, in principle, explain every function on the list, down to the last neuron, and still face a further question that the explanation never mentions. In Chalmers' words: "Why is the performance of these functions accompanied by experience?"
Two thought experiments made that gap famous. One is Frank Jackson's, from "Epiphenomenal Qualia" in the Philosophical Quarterly in 1982. Mary "is a brilliant scientist" who, for whatever reason, must investigate the world from a black-and-white room through a black-and-white television monitor. There she acquires "all the physical information there is to obtain about what goes on when we see ripe tomatoes, or the sky." Then she is let out. "Will she learn anything or not?" Jackson thought it obvious that she would, and drew the conclusion: "Ergo there is more to have than that, and Physicalism is false."
The other is the zombie. "Creatures that lack consciousness but are physically and functionally identical to ordinary human beings are called zombies," as the Internet Encyclopedia of Philosophy puts it. If such creatures are even conceivable, the argument runs, then the physical and functional facts leave the experiential ones open.
Philosophers take both seriously, and they split over them. The 2020 PhilPapers Survey, which Chalmers conducted with David Bourget, drew 1,785 respondents. Of those who answered the question on zombies, 24 percent accepted or leaned toward the view that zombies are metaphysically possible, 37 percent thought them conceivable but not metaphysically possible, and 16 percent thought them inconceivable.
Then comes the twist that the argument's fame tends to hide. In the late 1990s, Jackson changed his mind. According to the encyclopedia, he "now defends physicalism and rejects the knowledge argument," and rejects in particular the claim that Mary learns new truths when she leaves the room. In 2003 he wrote that he had ended up agreeing with the philosophers who held that Mary gains new abilities rather than learning new truths, adding: "for the life of me, I cannot see how we could have known they were right without going via representationalism." The most famous argument that experience is something extra was given up by the man who made it. The literature it started carried on without him.
The same survey shows the field's mood in numbers, and they do not settle into one picture. Asked whether there is a hard problem of consciousness, 62 percent of respondents accepted or leaned toward yes. Asked about the nature of consciousness, the largest group, 33 percent, accepted or leaned toward functionalism, roughly the view that a mental state is defined by the work it does rather than by what it is made of. Dualism followed at 22 percent and identity theory at 13. So a majority think the hard problem is real, while the most popular single view is the one on which getting the function right would be enough.
The survey's question about other minds is the one an AI builder should frame and hang on the wall. Asked which groups have at least some conscious members, 95 percent accepted or leaned toward adult humans, 89 percent cats, 65 percent fish and 35 percent flies. Only 3 percent accepted current AI systems, and 82 percent rejected them. But 39 percent accepted future AI systems, more than accepted flies.
Koch and Chalmers settled their bet on the night the first results came in from an unusual experiment. Called an adversarial collaboration, it was built to test two leading theories of consciousness against each other: integrated information theory, or IIT, and global neuronal workspace theory. The theories' proponents and a theory-neutral consortium agreed in advance on the design and on what each theory predicted, and preregistered both. Then 256 people viewed clearly visible stimuli while their brain activity was measured with fMRI, magnetoencephalography and intracranial EEG.
When the results were published in Nature in April 2025, the consortium's summary was blunt. The findings "align with some predictions of IIT and GNWT, while substantially challenging key tenets of both theories." For IIT, a lack of sustained synchronization within the posterior cortex "contradicts the claim that network connectivity specifies consciousness." The workspace theory was challenged by "the general lack of ignition at stimulus offset", the burst of activity it predicted when a stimulus disappears, and by the limited representation of some conscious content in the prefrontal cortex. On stage in 2023, Lucia Melloni, who helped lead the experiments, explained her neutrality another way. "My thought is that I come from a family of divorced parents," she said. "And you love them both."
The fight did not stay in the lab. On 15 September 2023, a letter titled "The Integrated Information Theory of Consciousness as Pseudoscience" appeared on the preprint server PsyArXiv. Its author list ran to 123 named researchers plus a collective signature, "IIT-Concerned." Among them were Daniel Dennett, Keith Frankish, Yoshua Bengio and Bernard Baars, whose global workspace theory Chalmers had analyzed back in 1995. The letter objected partly to the media coverage and partly to what the theory implies: "According to IIT, an inactive grid of connected logic gates that are not performing any useful computation can be conscious," it said, "possibly even more so than humans." Until the theory as a whole is testable, the authors wrote, "we feel that the pseudoscience label should indeed apply." The posted version notes that it was a short letter, then being revised into a longer article.
For people who build AI systems, the most useful part of this fight is where the theories disagree: over whether getting the function right is what counts.
In August 2023, nineteen researchers, with Patrick Butlin and Robert Long as joint first authors and Yoshua Bengio and Jonathan Birch among the rest, published a report on consciousness in artificial intelligence. They adopted a working assumption, computational functionalism: "Implementing computations of a certain kind is necessary and sufficient for consciousness, so it is possible in principle for non-organic artificial systems to be conscious." From recurrent processing theory, global workspace theory, higher-order theories, predictive processing and attention schema theory, they derived a list of "indicator properties" stated in computational terms, so that a system's architecture can be checked against it. Their conclusion was that "no current AI systems are conscious," but also that "there are no obvious technical barriers to building AI systems which satisfy these indicators."
They left one theory out and said why: "We do not consider integrated information theory, because it is not compatible with computational functionalism." IIT's proponents had already given its verdict on machines. In a 2015 paper, Giulio Tononi and the same Christof Koch who would pay off the bet wrote that, "in sharp contrast to widespread functionalist beliefs," IIT implies that digital computers, "even if their behaviour were to be functionally equivalent to ours, and even if they were to run faithful simulations of the human brain, would experience next to nothing." A purely feed-forward system that copied the brain's input-output behavior, they added, would be "certainly capable of passing the Turing test, yet it would have zero Φ and would thus be a 'perfect' zombie."
So the same machine gets opposite verdicts from two leading research programs, and the disagreement is less about evidence than about what to look at. On the indicator approach, the right architecture could make a system a serious candidate for consciousness. On IIT, as its proponents read it, no program on a conventional digital computer is likely to be conscious, and on its critics' reading a grid of idle logic gates might be. Six of the report's nineteen authors also signed the pseudoscience letter, and four of them were among the ten who drafted it. The fault line in the science runs where Chalmers drew his line in 1995.
That leaves anyone who builds, buys or regulates AI systems with a practical rule. When a claim about machine consciousness reaches you, from a vendor, a headline, a user or your own model's transcript, the first question is: on which theory? A claim that does not say is not yet a claim, and that cuts both ways. It said it was afraid, so it feels assumes a theory. So does it is only predicting the next token, so it can't. Each answer takes a side on the question Chalmers posed, and each is worth exactly as much as the theory it assumes.
Two habits follow. One is to stop treating behavior as evidence about experience. The indicator report calls behavioral tests unreliable, because "AI systems can be trained to mimic human behaviours while working in very different ways." In Chalmers' terms, a system that reports a feeling is performing reportability, one of the easy problems, and explaining that function is the easy part. The other is to keep two ledgers. In your model cards, policies and product copy, separate what a system does, which you can measure, from what it might experience, which depends on a theory, and never let the first silently answer the second.
Koch and Chalmers will settle their second bet in 2048, if both are still around. Until then, the honest answer about any machine starts with a theory and a probability, and Chalmers, who suspects he will collect again, says he hopes to lose.
Sources
How this was checked. Every quotation was checked word for word against the saved text of the source named beside it. Jackson's 1982 wording was read in a scan of the journal pages, and his 2003 sentence is quoted as the encyclopedia quotes it. The survey figures were read from the results page. The count of 123 named researchers comes from the letter's posted author list, whose first entry is the collective "IIT-Concerned". The six authors shared by the report and the letter were found by matching surname and first initial across the two author lists.
Chain of Consciousness keeps a tamper-evident, dated record in hash-linked entries of what an AI system actually did, so the behavior you can measure stays checkable, separate from any claim about what the system experiences.
pip install chain-of-consciousness npm install chain-of-consciousness
Or start without installing anything: Hosted Chain of Consciousness.