The Chinese room
A man in a room follows a rulebook to answer questions in Chinese perfectly, without understanding a word. Is that all a computer ever does?
- Reading time
- 4 min
- Difficulty
- Intermediate
- Branch
- Mind
- First posed
- John Searle, 'Minds, Brains, and Programs', Behavioral and Brain Sciences, 1980
The setup
Imagine you are locked in a room. You do not speak or read any Chinese. Through a slot, people pass you cards covered in Chinese characters. You have a huge rulebook, written in English, that tells you exactly which characters to send back for any characters you receive. The rules refer only to the shapes of the symbols: 'if you see this squiggle, followed by that squoggle, write this other squiggle'.
You follow the rules carefully. Outside the room, native Chinese speakers are delighted. The answers coming out of the slot are fluent, witty and appropriate. As far as they can tell, whoever is in the room understands Chinese perfectly. The room passes the Turing test in Chinese.
But you understand nothing. You have no idea that you have just been asked about your favourite food, or that you replied with a joke about dumplings. You are manipulating symbols by their shape alone.
John Searle published this thought experiment in 1980. His point: a computer running a program is in exactly your position. It manipulates symbols according to formal rules. So however well it performs, running a program cannot by itself produce understanding. In Searle's slogan: Syntax is not sufficient for semantics.
Why it matters
The Chinese room is perhaps the most widely discussed argument in the philosophy of artificial intelligence. With modern language models producing fluent text, it has become more relevant, not less. Is a system that predicts the next word in a sentence, without any sensory contact with the world, in the position of the person in the room?
It also raises a deep question about our own minds: what is understanding, and what would it take for a physical system, a brain or a computer, to have it?
Key terms
- SyntaxThe formal arrangement of symbols according to rules, regardless of what they mean. Searle argued that computers operate on syntax alone.
- SemanticsMeaning: what words, sentences or symbols are about or stand for. Searle contrasts it with syntax.
- Strong AISearle's term for the claim that a suitably programmed computer would literally have a mind and understand, rather than merely simulate understanding.
- FunctionalismThe view that mental states are defined by what they do, their causal roles linking inputs, other mental states and behaviour, rather than by what they are made of.
- Symbol grounding problemThe problem of how symbols in a system, such as words in a computer, come to mean anything, rather than just being linked to other meaningless symbols.
The positions
Searle is right
Running a program, however sophisticated, is not sufficient for understanding. Minds need more than formal symbol manipulation.
The systems reply
The person does not understand Chinese, but the whole system (person, rulebook, room) does.
The robot reply
Put the program in a robot with cameras and arms, so its symbols connect to the world, and it would understand.
The intuition is unreliable
The thought experiment pumps a misleading intuition. A system complex enough to pass the test would be so vast and fast that our intuitions about a man shuffling cards do not apply.
Arguments and objections
Searle is right
- Best argument
- The thought experiment shows that one can run the program without understanding. The person in the room is doing everything a computer does. If they do not understand, nothing in the program does. Understanding requires the causal powers of brains, or something equivalent, not just the right input-output behaviour. Searle did not say machines could never think; brains are machines. He said programs alone could not.
- Strongest objection
- Critics say Searle never explains which causal powers brains have that computers lack, or why those powers could not be implemented in silicon. The claim that meaning requires biology looks like an unsupported assumption.
The systems reply
- Best argument
- The person is like a single neuron, or the central processor in a computer. No neuron understands English, but you do. The understanding belongs to the system as a whole, not to any one component. Searle's argument asks whether a part understands and concludes that the whole does not, which is a fallacy.
- Strongest objection
- Searle answered that he could memorise the entire rulebook and do all the work in his head, walking around outside. Now he is the whole system, and he still understands no Chinese.
- The reply
- Systems theorists reply that, if Searle memorised the rules, there would be two systems running on one brain: Searle, who speaks English, and a Chinese-speaking system that Searle has no access to, much as in some cases of dissociation.
The robot reply
- Best argument
- Understanding requires that symbols be grounded in perception and action: 'tree' means something because it connects to seeing and touching trees. Stevan Harnad called this the symbol grounding problem problem. The room is sealed off from the world. A robot that learned words by interacting with things would not be.
- Strongest objection
- Searle answered that sensor inputs are just more symbols in the slot, and motor outputs just more symbols out of it. The man in the room would still be manipulating shapes without knowing they came from cameras.
The intuition is unreliable
- Best argument
- Paul and Patricia Churchland compared Searle's argument to saying that waving a magnet in a dark room produces no visible light, so light is not electromagnetic radiation. The slowness and scale of the imagined system mislead us. Real understanding may emerge from enormous numbers of simple operations, as it does in brains.
- Strongest objection
- Searle's defenders reply that speed and scale are irrelevant to the logical point: symbol manipulation by shape, however fast, does not involve meaning.
Who argued what
- Searle is right John Searle, Hubert Dreyfus
- The systems reply Daniel Dennett, Ned Block, Jack Copeland
- The robot reply Stevan Harnad, Jerry Fodor
- The intuition is unreliable Paul Churchland, Patricia Churchland, Daniel Dennett
Variations
The language model in the room
A modern language model is trained on vast amounts of text and learns to predict which words come next. It has never seen a tree or tasted food. Is it the man in the room, with a statistical rulebook instead of a written one? Or does learning the deep structure of language amount to some kind of understanding?
The brain simulator
Suppose the rulebook simulates the firing of every neuron in a Chinese speaker's brain. Searle says the man in the room still understands nothing. Critics ask: if a perfect simulation of a brain would not understand, what is special about the brain?
Go deeper
Strong and weak AI
Searle distinguished 'weak AI', the claim that computers are useful tools for studying the mind, from 'Strong AI', the claim that an appropriately programmed computer literally has a mind. He accepted weak AI entirely. His target was only the claim that running a program is sufficient for understanding.
Related ideas
Further reading and sources
- EncyclopediaStanford Encyclopedia of Philosophy: The Chinese Room Argument
- WikipediaChinese room
- BookMinds, Brains and Science, John Searle
- PaperMinds, Brains, and Programs, John Searle
Open questions
- What would convince you that a machine understood you?
- Does your own brain do anything more than follow rules?
- Can meaning exist without experience?