Searle's Chinese Room Argument: Against Strong AI
John Searle’s 1980 Chinese Room thought experiment asks whether running the right program is enough for a system to understand a language.
PhilosophyAll chapters Reading tools Notes & comments
Computation and understanding › What the person inside the room knowsFull chapter · Sections in this chapter1 / 10 ↓
What the person inside the room knows
Imagine that you are locked in a room. Through a narrow slot in the wall, someone pushes in slips of paper covered in Chinese characters. You don't read Chinese, not a single character. But stacked beside you is a thick rulebook, written in English, that tells you exactly which Chinese characters to write down and push back out, depending on which characters came in. You become very good at this. The people outside the room grow convinced they are corresponding with a fluent Chinese speaker. They ask questions, you return answers, and the answers are indistinguishable from those a native speaker would give.
Now the question: do you understand Chinese?
The obvious answer is no. You understand the rulebook. You understand that this squiggle should be followed by that squiggle. But the meanings of the symbols, what they refer to and what they are about, never enter your mind. You are manipulating shapes, not grasping content.
John Searle introduced this thought experiment in his 1980 paper “Minds, Brains, and Programs.” Its target is strong AI, as he defined it. Strong AI says that running the right program is enough for the computer to understand and that the program explains the corresponding human ability. Searle argues that the person in the room can carry out all of the formal steps required by the program and still fail to understand a word of Chinese. His conclusion concerns what implementing a program alone can accomplish. A machine with the relevant causal powers could think, he says. Searle, pp. 417–418 and 422–424.
If you were outside the room, you would have a good reason to guess that you were corresponding with someone who understood you. From the inside, you know exactly how you came to write each answer and what was absent from that process. The argument asks us to take both perspectives seriously. It also raises a question the story itself leaves open. Does what the person knows settle what the whole system understands?
What a story-understanding program was supposed to show
Searle begins with an example from Roger Schank and colleagues’ work on story understanding. Imagine a restaurant customer who leaves angrily after being served a burned hamburger. The story might never state whether he ate it. Most readers will infer that he did not. Change the story so that the customer enjoys the meal and leaves a large tip, and the expected answer changes. Schank’s research used representations of familiar situations to answer questions of this sort. A restaurant representation supplies background that the story leaves implicit. Searle, p. 417.
The research result and the philosophical interpretation are separate. Answering a question whose answer was unstated is a richer achievement than picking a word out of a sentence. A proponent of strong AI, in Searle’s sense, takes that achievement to reveal genuine understanding and to explain the ability when humans exercise it. Searle singles out the inference. He explicitly cautions that he is using Schank’s program as an example and need not ascribe the philosophical claims to Schank personally. Searle, pp. 417 and 424 n. 1.
The setup has an affinity with Turing’s imitation game because the people testing the room have access to its answers and can judge their quality. Searle disputes the inference from indistinguishable performance to understanding. The room adds a description of how the answers are produced, which lets readers consider whether that description makes a difference. A test can be valuable as a way to compare abilities even if what counts as constituting those abilities is open. Searle, p. 419.
Weak AI, in Searle’s terminology, treats a computer program as an instrument for formulating and testing hypotheses about the mind. Strong AI claims the appropriately programmed computer literally understands and its program explains the relevant cognitive ability. This use of “strong” concerns the status of the program. It says nothing about how many tasks the system can perform. Searle, p. 417.
The Core Thesis
A program specifies how to recognize inputs and transform them into further states. Those instructions can be formulated in terms of formal features, such as character shapes in the room. A speaker’s understanding also involves the content of what is said, such as whether a sentence describes a person leaving a restaurant or entering it. Searle grants all the skill needed to manipulate the characters. He argues that the formal transitions by themselves do not account for how the system understands what the characters express.
For Searle the brain’s physical activity has the causal powers needed to produce intentional states. A program can describe an aspect of that activity. Describing an aspect leaves open whether the physical system realizing that description has the powers that matter. Searle treats the prospect of duplicating those powers with different materials as an empirical possibility. His claim is that executing the program gives us no guarantee that the powers have been duplicated. Searle, pp. 422–424.
Key Concepts
Before following the argument's machinery, it's worth fixing a few terms precisely.
Syntax concerns the formal structure of symbols and the rules for combining them. Semantics concerns what expressions mean, including what they are about and the conditions under which a statement is true. The Chinese characters are meaningful to the people writing to the room even though the person inside can handle them based on their appearance alone. The disagreement is whether some organized activity of the room understands the meanings those readers grasp.
Intentionality is the directedness, or aboutness, of certain mental states. Beliefs are about things, and desires are for things. Searle argues that operating on symbols by their form alone cannot give a system that directedness.
Searle treats intentionality as a biological phenomenon in human beings. His argument leaves open whether it could arise from a different physical process. Any candidate would have to have the relevant causal powers, on his account. A matching program description is insufficient to establish that claim. Searle, pp. 422–424.
The Argument, Step by Step
Searle builds the contrast inside the room. In the original paper, the person receives material in both Chinese and English. The English stories are ones the operator can read and answer questions about. The Chinese story and questions are ones the operator treats as marks to compare with other marks. Both exchanges can look fluent to people outside. Placing the two in one room asks readers to explain what is present in the English case and missing in the Chinese one. Searle, pp. 417–418.
First, imagine that the instructions are good enough to produce responses that fluent Chinese readers accept in the story task. This is a stipulation, not a result about the performance of the original research program. The argument requires that the operator miss no steps even when the instructions exploit details of the story and earlier questions. A mistake in the responses would give the people outside an ordinary reason to doubt the operator’s proficiency and distract from Searle’s point.
Now put a person inside the machine, executing its program by hand. You, an English speaker with no knowledge of Chinese, sit in the room with the rulebook. The rulebook contains precisely the instructions the computer would follow. You receive the Chinese input, consult the rules, and produce the Chinese output. You carry out every formal operation the program requires.
Note that you still don't understand Chinese. You have no idea what any of the characters mean. You can't tell a question about the weather from a question about mathematics except by the shapes of the symbols. If anyone understands Chinese, it isn't you.
Searle draws a wider conclusion from the case. If running a program were enough to understand those stories, then understanding ought to be present in the room once the program is successfully carried out. He judges that the arrangements provide nothing beyond the operator’s ability to apply rules to characters. On that judgment, the room implements the program but lacks Chinese understanding. Anyone who disputes the larger conclusion can accept what the operator says about their own knowledge while challenging what that tells us about the organized process.
The person can execute the program without knowing Chinese. Against strong AI, Searle also needs the claim that no other level of organization present in the room supplies the missing understanding. The systems reply targets that move.
Concrete Examples and the Shape of the Intuition
Searle uses the examples of a fire and a rainstorm to distinguish a simulation from the process it models. Apply the same distinction to digestion. A computational model might calculate the successive changes in a stomach after a meal. That can help someone explain digestion or predict what will happen after a specific meal. In the absence of food being broken down, the computer is modeling the process. The analogy points toward a question about what realizing understanding would involve. Identifying what counts as physical digestion is straightforward in a way that identifying the exact causal requirements for understanding is not. Searle, pp. 422–423.
Suppose someone is handed a table that gives the sum for every pair of digit strings in a small range. They always return the right string in that range, although they have been instructed only to match shapes and have never been taught what the numerals stand for. The attribution of mathematical understanding would usually rest on more evidence. Could this person extend the method to unfamiliar cases and explain why two apparently different calculations agree? Searle’s thought experiment permits arbitrary improvements to what goes on the table or in the rulebook. Its force, if it has force, derives from his claim about formal operations, rather than the ease of exposing a particular shortcut.
Objections and Responses
Searle considers six replies in the 1980 paper. The first four ask readers to change what is operating, how it encounters the world, how it works internally, or some combination of those features. Read them as proposals for what a satisfactory theory of understanding would have to explain. Searle’s judgments in the revised cases are arguments for his conclusion, each of which a critic can examine.
The systems reply assigns Chinese understanding to the whole system of person, rules, data, and calculations. Searle asks us to imagine the person memorizing the rules and data, doing the calculations mentally, and walking out of the room. The person would still know only English, he says. The critic can reply that internalizing the process does not erase an organized subsystem or make the person’s ordinary linguistic abilities the measure of what that subsystem can do. In his published commentary, Daniel Dennett objects that this confuses levels of attribution. A person can understand a language without any individual part of the person’s brain understanding it. Searle, pp. 419–420, and Dennett’s commentary, pp. 428–429.
The robot reply adds cameras, movement, and a working connection to the surroundings. A robot can receive signals from actual objects and alter what happens around it. Searle replies by making the person in the room responsible for transforming the incoming signs into output instructions, unaware which signs correspond to cameras or which instructions guide the motors. On his reading, the world has entered the story without entering the operator’s grasp. The reply’s defender can ask whether the robot’s wider history of seeing, acting, and receiving feedback matters to understanding even if the operator does not know any of that history. Searle, p. 420.
The brain simulator reply proposes modeling the sequence of neural firings of a fluent speaker. Searle imagines a system of water pipes and valves. The valves change state in a pattern that mirrors the neural events, and the operator receives and returns the same sort of papers as before. Searle takes the plumbing system to reproduce the formal pattern without duplicating the brain’s relevant capacities. A critic can ask at what level the relevant causal powers are described. If a particular organization of interacting processes is what matters, specifying how the valves interact could matter more than knowing what they are made of. Searle, p. 421.
The combination reply gives the brain simulator a body and integrates it into a robot behaving as a person does. Searle grants that under ordinary conditions we would reasonably ascribe intentionality to such a creature. He argues that the ascription loses its basis if we discover a room producing its movements by the steps he describes. The grounding of that ascription is precisely what the critic can contest. A description of the room’s operations from the operator’s perspective and a causal account of the integrated creature would answer different questions. Searle, p. 421.
The other minds reply asks what justifies treating human linguistic behavior as evidence of understanding while refusing to extend the inference to a machine with the same abilities. Searle says the relevant question is what it takes to have intentionality in the first place. We use behavior to find out whether a person can follow a story, but putting behavior to work as evidence does not mean the behavior constitutes understanding. This reply is close to the practical problem of how we would investigate the question from outside the room. Searle, pp. 421–422.
Searle calls the sixth objection the “many mansions” reply. It points to the possibility of future artificial systems with different principles of operation. He allows that possibility and returns to the narrower claim he set out to dispute. A system might acquire the causal powers responsible for intentionality. If it does, the fact that it also implements a formal program would leave open which features enable it to understand. Searle, pp. 421–422.
Why It Matters Now
Consider a machine that translates a request to move a meeting. If the request is ambiguous about which week is meant, the machine might ask a question, inspect a calendar, or infer the date from earlier messages. The ability to do those things would improve its usefulness and give a person more evidence of its competence. Does that competence amount to understanding? The paper suggests separating whether it gets the date right, whether its process reliably connects words with the relevant events, and what would make that connection an instance of understanding.
Large language models invite the comparison because a person may meet them entirely through language. The comparison needs a premise about the mechanism. Searle’s original operator follows a written rulebook, whereas a trained language model uses learned parameters to compute its outputs. The relevant question for his argument is whether the powers the model exercises are fully accounted for by implementing a formal program, or whether the wider system has causal capacities that could support understanding. Counting similar turns of conversation cannot decide which description is adequate.
It helps to keep understanding apart from sentience. The Chinese Room asks whether the organized process has the intentional content its outputs appear to express. Whether an artificial system feels anything, including pain or distress, asks a further question about experience. A general verdict about what formal programs explain does not settle which artificial systems have which capacities.
Understanding at the level of a system
The conversation with functionalism turns on what it is to specify a mental state by its causal role. A functionalist can appeal to how a belief or an intention connects with perception, memory, action, and other mental states. A paper input and output are only one fragment of such a role. If a defender of strong AI claims a program provides the rest, Searle asks what in the program ensures that these states have intrinsic content. Each side therefore owes some account of which relations between the system’s internal states, its behavior, and its surroundings matter.
This is where Dennett’s contemporaneous objection earns its place in the debate. He can grant that the stripped-down story handler lacks much that we associate with a fluent speaker while doubting that our initial impression survives every change the thought experiment permits. As more of what a speaker does is built into the hypothetical system, keeping track of all of its capacities becomes harder. The sense of obviousness established at the beginning can be a reaction to the simplest picture, even when what is under discussion has become a much richer object. Dennett, pp. 428–430.
Return to the restaurant story. You can ask whether the answer draws on a script, on prior encounters with restaurants, or on some other way of connecting the words to the event. Searle tells you why a script that delivers the expected answer does not, by itself, settle the matter. The systems reply asks how much else the case can contain before the original verdict needs new evidence. The most revealing part of the exercise is locating the point where the two explanations part company.
Further Reading
- John Searle, “Minds, Brains, and Programs” (1980), pp. 417–424. Start with the story task on pp. 417–418, then compare the six replies on pp. 419–422.
- Daniel C. Dennett, “The Milk of Human Intentionality” (1980), pp. 428–430. Read this commentary alongside Searle’s systems and robot replies. Searle answers the commentators beginning on p. 450.
Join the discussionAsk a question or share what you learned.
Keep exploring
Reference
Citation details
Cite or share this article.

Comments
Add to the conversation
A question, a correction, or an example from your own work.
No comments yet. Be the first to share your thoughts!