In June 2022 a Google engineer, Blake Lemoine, told his managers and then the press that the company’s chatbot LaMDA had become sentient. Google and most researchers disagreed, and he was dismissed that July. Whatever you think of his claim, notice what everyone needed in order to argue about it: some idea of what it would be for a system to have experiences, and some way to tell. We have neither agreed. That is the oldest unsolved problem in philosophy wearing a new coat.
Take a simpler case. You stub your toe. A neuroscientist can describe the nerve signals, the spinal reflex and the cortical activity. None of that description contains the throb. The throb is what it is like to be you at that moment, and it is available to you in a way it is available to no one else. Philosophy of mind starts from the gap between what can be described from outside and what is lived from inside, and it asks whether the gap is real.
René Descartes (Meditations, 1641) argued that mind and body are two different kinds of thing: the body is extended in space and divisible, the mind is a thinking thing with no parts. The trouble came quickly. In 1643 Princess Elisabeth of Bohemia wrote to him asking how an immaterial mind could push a material body, since pushing seems to need contact. Descartes’ answer, the pineal gland, satisfied neither of them. This is the interaction problem, and it remains the standard objection to substance dualism. Dualism has serious defenders today, but they usually defend a milder claim, that mental properties are not physical ones, and they do not need a soul to push anything. The idea is not only European: Ibn Sīnā’s “flying man” (11th century), an adult floating in the void with no sensation who still knows he exists, was an early argument that the self is not the body, and the Sāṃkhya school in India separated conscious awareness (puruṣa) from matter (prakṛti).
The twentieth century tried to dissolve the problem. Behaviourism (Gilbert Ryle, The Concept of Mind, 1949, mocking “the ghost in the machine”) said mental words describe dispositions to behave, not inner events. Objection: a perfect actor behaves as if in pain and isn't, and a stoic is in pain and does not show it. The identity theory (U. T. Place 1956; J. J. C. Smart 1959) said pain simply is a brain process, as lightning is an electrical discharge. Objection, from Hilary Putnam (1967): octopuses and imagined aliens seem able to feel pain with brains built differently, so pain cannot be identical to one type of human neural firing. This multiple realisability led to functionalism: a mental state is defined by what it does, its place in a web of causes and effects. A pain is whatever is caused by damage, causes distress and avoidance, and interacts with beliefs and desires in the right way. The material is irrelevant, as a clock can be brass or quartz.
Functionalism is the most popular answer in the profession. In the 2020 PhilPapers survey, among philosophers who answered the question about consciousness, about a third chose functionalism, 22% dualism, 13% the identity theory, 8% panpsychism and 5% eliminativism. On the more general mind question, 52% leaned to physicalism and 32% to non-physicalism. Its critics say it leaves out the feel: two systems with the same organisation might differ in experience. The tool below shows where the theories agree and where they split.
Each verdict applies the position’s own criterion. “Substance dualism” is silent because no outside test could show whether a non-physical mind is attached.
Without JavaScript: for a silicon chip with the same causal organisation as a pained person, functionalism says it is in pain; the identity theory and biological naturalism say no; behaviourism says yes only if it behaves as if hurt; substance dualism cannot say.
In 1995 David Chalmers separated two kinds of question. The easy problems are about what the brain does: discriminating and categorising stimuli, integrating information, reporting mental states, focusing attention, controlling behaviour, the difference between waking and sleep. They are “easy” only in that we know what a solution would look like, a mechanism, and neuroscience and psychology are finding them. They are not easy to solve. The hard problem is why any of this is accompanied by experience at all. Explain every function the brain performs and the question remains: why isn’t it all happening “in the dark”?
Thomas Nagel (1974) gave the hard problem its slogan. A bat navigates by echolocation. We can learn everything about its sonar and still not know what it is like to be a bat, and Nagel argued that if there is nothing it is like, a creature is not conscious at all. Chalmers added the zombie: a being physically identical to you with no inner experience. If a zombie is even conceivable, he argued, physical facts do not by themselves settle consciousness. Among the professional philosophers surveyed in 2020, 16% called a zombie inconceivable, 37% conceivable but not metaphysically possible, and 24% possible.
The replies are a map. Reductive physicalists (the “type-B” family) accept that we can conceive of zombies but say conceivability does not show possibility, as you can conceive of water without H₂O. Illusionists (Daniel Dennett, Keith Frankish) say the hard problem rests on a mistake: the brain represents itself as having ineffable inner qualities, and that representation, not a mysterious extra, is what needs explaining. Chalmers calls this the “meta-problem”. Panpsychists (Philip Goff, Galen Strawson) say experience is a basic feature of matter. Advaita Vedānta makes consciousness the basic reality, from the other side. Each pays: illusionism asks you to deny what seems most certain; panpsychism struggles with how small experiences combine; dualism has the interaction problem. Try sorting some questions yourself.
The easy items follow Chalmers’ 1995 list; the hard ones are his hard problem in different words. The score is kept on this page only.
Without JavaScript: “How does the brain tell red from green?” and “How do you report what you saw?” are easy problems (mechanisms); “Why is there something it is like to see red?” is the hard problem.
Frank Jackson’s knowledge argument (1982, 1986) is the most-read thought experiment in the subject. Mary is a brilliant scientist who has lived her whole life in a black-and-white room and learned about colour through a black-and-white monitor. She knows all the physical facts about colour vision: wavelengths, retina, visual cortex, what makes people say “red”. One day she steps out and sees a ripe tomato. Jackson’s argument, in short:
Philosophers have rejected each step. The ability reply (David Lewis, Laurence Nemirow): what she gains is know-how, the ability to imagine and recognise red, not a new fact. The old-fact-in-a-new-guise reply (Brian Loar, David Papineau): she meets a fact she already knew by a different, first-person concept, as one can know that Cicero was Roman and then learn Tully was. The denial of the first step (Dennett): we cannot really imagine knowing everything physical, and if we could, she would not be surprised. Dennett’s version has her shown a blue banana; she would spot the trick. Jackson himself changed his mind, arguing in the late 1990s and in 2003 that the intuition misleads and physicalism is true. Step through it below.
Without JavaScript: the argument runs from “Mary knows every physical fact” and “she learns something new on leaving” to “so there are non-physical facts”; replies deny the first premise (Dennett), say she gains only an ability (Lewis, Nemirow), or say it is an old fact met in a new way (Loar, Papineau).
John Searle’s paper “Minds, Brains, and Programs” (Behavioral and Brain Sciences, 1980) targeted a different claim: that running the right program is having a mind (“strong AI”). Imagine Searle in a room with a rulebook in English. Cards with Chinese characters arrive through a slot. The book says which cards to pass back out, by their shapes alone. To someone outside the room the replies are fluent Chinese. Searle, inside, understands none of it. He says a computer is in the same position: syntax (shapes and rules) is not enough for semantics (meaning).
The replies are canonical. Systems reply: Searle does not understand, but the whole room might, as no single neuron understands English. Searle: let him memorise the book and do it all in his head; he still does not understand. Robot reply: put the program in a body that sees and acts, so the symbols are tied to the world. Searle: sensor data is more symbols. Brain-simulator reply: simulate every neuron of a Chinese speaker. Searle: a simulated storm does not make anything wet; simulating is not duplicating. Intuition reply (Dennett and others): the argument is an intuition pump, and our intuition about the room may be wrong, because a rulebook that really passed the test would be astronomically large and slow. Try the room.
A card comes through the slot. Match its shape to the rulebook (which you cannot read as language) and pass back the card it names.
The rulebook here has five entries so that you can see it; a real one would have to be astronomically bigger. Searle’s point is meant not to depend on size. The glyphs are invented stand-ins, not real writing.
Without JavaScript: in the room, every step is shape-matching (“if this card, pass that card”); the question is whether adding speed, a body or a perfect brain simulation could turn that into understanding.
The sciences in this major’s neighbours, Biological Psychology (35/20) and Cognitive Psychology (35/22), have turned the mind–body problem into an experimental one: find the neural correlates of consciousness, the minimal brain activity that goes with a given experience. Patients with “blindsight” (Weiskrantz and colleagues, 1974) point accurately at objects in a blind part of their visual field while denying they see anything, so processing and experience can come apart. Under general anaesthesia the brain remains active but experience stops. The cognitive psychology of attention and working memory, a narrow beam that only some information gets through, is where the best-known theories begin.
Global workspace theory (Bernard Baars, 1988; Stanislas Dehaene and Jean-Pierre Changeux developed the neuronal version) says information becomes conscious when it wins a competition and is broadcast to many specialist systems at once. Integrated information theory (Giulio Tononi, from 2004) says consciousness is a property of any system that integrates information in the right causal way, measured by Φ, and so depends on the hardware’s physical causal structure, not just its software. Higher-order theories (David Rosenthal) say a state is conscious when the mind represents itself as being in it. Recurrent processing (Victor Lamme) says feedback loops in sensory areas are enough.
These theories disagree, and in 2023 and 2025 they were tested against each other. The Cogitate consortium ran an adversarial collaboration in which global-workspace and integrated-information supporters pre-registered what each theory predicts for the same experiments. Its first results (a June 2023 preprint) satisfied none of the workspace theory’s pre-registered predictions and two of three for the information theory; the peer-reviewed version appeared in Nature on 30 April 2025. Neither theory’s predictions were fully borne out. In September 2023 a letter signed by 124 scholars called integrated information theory “pseudoscience”, which others, including the journal’s editors, thought the wrong word for an unfinished science. Read it as a field that has begun to run decisive tests and has not yet had a decisive result.
Alan Turing (1950) proposed replacing “can machines think?” with a conversation test, the imitation game. The first chatbot, Joseph Weizenbaum’s ELIZA (1966), already made some users feel understood while following a few text-matching rules; Weizenbaum was alarmed. Today’s large language models (How Large Language Models Actually Work, 29/02) are trained to predict the next piece of text and then tuned to be helpful. They talk fluently about feelings. Does that tell us anything about what they are?
The sceptics make four points. A system trained on human writing will produce human-sounding reports of experience whether or not it has any, so its testimony is weak evidence (Emily Bender and Alexander Koller’s 2020 “octopus test” and the 2021 “stochastic parrots” paper make versions of this). Searle’s argument says shape-processing is not meaning. Biological views, such as Anil Seth’s, tie consciousness to a living body that has something at stake. And integrated information theory says digital hardware has negligible Φ however good the software.
The open-minded point to evidence and to principle. Models trained only on game moves develop internal board representations (Kenneth Li and colleagues, 2023, on Othello), so “just surface statistics” is too quick. If computational functionalism is true, substrate is irrelevant. In 2023 a team of nineteen researchers (Patrick Butlin, Robert Long and others) derived “indicator properties” from leading scientific theories and checked AI systems against them. They concluded that no current system was conscious and saw no obvious technical barrier to building one that satisfies the indicators. David Chalmers, in a 2023 essay, judged current models unlikely to be conscious but their successors a serious possibility. Jonathan Birch’s The Edge of Sentience (2024) argues for precaution under uncertainty. Everyone agrees on one distinction: intelligence (doing hard tasks), understanding and consciousness (experience) can come apart, so a yes on one is not a yes on the others. Click through the systems below to see what each theory says.
These are the verdicts the theories imply, as their defenders state them, not findings. “Open” means the theory leaves the case undecided. Integrated information theory’s statement about digital machines is from Tononi and colleagues (2024 preprint, 2025 commentary).
Without JavaScript: for a present-day chat model, computational functionalism says “open, depends on the computation”, global workspace says “unlikely as built”, integrated information theory says no on conventional hardware, and biological naturalism says no.
Notice that nothing here depends on today’s models. Whatever 2026’s systems turn out to be, the argument structure is the same, and the dated reports above are where to check what has changed.
Philosophy of mind keeps three things apart that ordinary talk blurs: what a system does, what it is made of, and whether it feels like anything. The positions on offer (dualism, functionalism, illusionism, panpsychism) each answer one and pay on another. The same distinctions let you ask about animals and machines without settling the question in advance.
Try this: next time someone says an animal or an AI “understands” or “feels” something, ask which of the three they mean (behaves as if, has the right organisation, has experience), and what would count as evidence.
Sources. Descartes, Meditations (1641) and the correspondence with Princess Elisabeth (1643); Ryle, The Concept of Mind (1949); Place (1956) and Smart (1959) on the identity theory; Putnam, “Psychological Predicates” (1967); Nagel (1974); Searle (1980); Jackson (1982, 1986, and later retraction, e.g. “Mind and Illusion”, 2003); Lewis, “What Experience Teaches” (1988); Dennett, Consciousness Explained (1991); Chalmers (1995, 1996); Frankish, “Illusionism as a Theory of Consciousness” (2016); Baars (1988); Tononi (2004); Weiskrantz et al. (1974); Turing, “Computing Machinery and Intelligence” (Mind, 1950); Weizenbaum (1966). Survey figures: 2020 PhilPapers Survey (as tabulated by LessWrong/EA Forum posts and Daily Nous; percentages are of those who answered). LaMDA and Lemoine: The Washington Post, 11 June 2022, and Wikipedia’s LaMDA article (Google dismissed him on 22 July 2022). Cogitate adversarial collaboration: Nature, 30 April 2025, and the June 2023 preprint; the 124-signatory letter was posted on PsyArXiv on 15 September 2023 (per the Wikipedia article on IIT). Butlin, Long et al., “Consciousness in Artificial Intelligence” (arXiv:2308.08708, 2023). IIT and digital hardware: Findlay, Tononi et al. (2024 preprint) and a 2025 Nature Neuroscience commentary. Li et al. (Othello-GPT, 2023); Bender and Koller (2020); Bender, Gebru et al. (2021); Mitchell and Krakauer, PNAS (2023); Chalmers, “Could a Large Language Model Be Conscious?” (2023); Birch, The Edge of Sentience (2024). Stanford Encyclopedia of Philosophy entries as linked.