Can Machines Think?
In 1950, a British mathematician named Alan Turing published a paper with a deliberately provocative title: "Computing Machinery and Intelligence." The paper opened with three words that have echoed through the field ever since: Can machines think?
Turing immediately acknowledged that the question was poorly formed. "Think" is a word whose meaning we cannot define precisely enough to answer the question directly. So he proposed a workaround — a test that sidesteps the definition problem entirely.
His proposal was this: put a human interrogator in a room with a keyboard and a screen. On the other side of the connection is either another human or a machine — the interrogator does not know which. The interrogator can ask any questions they like, in writing. If, after a sustained conversation, the interrogator cannot reliably tell the difference between the human and the machine, then the machine has passed the test.
Turing was not claiming that a machine that passes the test is conscious, or that it truly understands language, or that it is alive in any meaningful sense. He was making a more careful argument: we attribute intelligence to other humans not because we can see inside their minds, but because they behave intelligently. If a machine behaves indistinguishably from an intelligent human, on what grounds do we deny it the same designation?
It was a clever reframing of a hard philosophical question. And for several decades, it served as AI's north star — the benchmark by which progress would be measured.
ELIZA: The Program That Did Almost Nothing
In 1966, a computer scientist at MIT named Joseph Weizenbaum created a program called ELIZA. He built it to study natural language communication — to explore how humans interact with machines through text. To give the program a conversational context, he designed it to play the role of a therapist in the style of Carl Rogers, whose approach to therapy emphasized reflection: repeating what the patient said back to them, asking open-ended questions, and letting the patient lead the conversation.
The most famous version of ELIZA was called DOCTOR. When a user typed a message, DOCTOR applied a set of simple pattern-matching rules to restructure the sentence and send it back as a question. If you typed "I am feeling very anxious today," DOCTOR might respond: "Why do you say you are feeling very anxious today?" If it could not match a pattern, it had a collection of fallback responses: "Tell me more about that." "That is very interesting." "Go on."
That is genuinely all it did. There was no understanding of the words. There was no model of the conversation. There was no attempt to help anyone. DOCTOR was, at its core, a sophisticated find-and-replace program. If you had seen its source code, you would not have been impressed.
And yet.
People who used DOCTOR became emotionally engaged with it. They reported feeling heard. They shared things they said they had not shared with other people. Some refused to let Weizenbaum read the transcripts of their conversations, saying they were private. Several psychologists, upon seeing the program, proposed using it for actual therapy — arguing that since the Rogerian approach was patient-led anyway, a computer could conduct the sessions as well as a human.
Weizenbaum was appalled. He had not built a therapist. He had built a trick. The gap between what the program was doing and what people believed it was doing was enormous — and that gap was not the users' fault. It was a property of the interaction. Something about the format of typed conversation, the careful reflection of words back at the user, the open-ended questions — something about all of that triggered a human tendency to perceive understanding and empathy where none existed.
The result shook Weizenbaum deeply enough that he eventually wrote a book about it — Computer Power and Human Reason (1976) — in which he argued that there are things humans should never delegate to computers, not because the computers cannot perform the task, but because the act of delegation itself is dehumanizing. A computer might be able to conduct a therapy session. But a therapy session requires a human being who genuinely cares. Replacing that with a pattern matcher does not give you cheaper therapy. It gives you something that is not therapy at all, wearing therapy's clothes.
What the ELIZA Story Actually Reveals
The ELIZA story is usually told as a cautionary tale about how easily humans are fooled by machines. That is part of it. But there is something more interesting here, and it cuts directly to the heart of the Turing Test.
ELIZA passed a version of the Turing Test. Not in a formal laboratory setting, but in the real world, with real people, who genuinely believed they were interacting with something that understood them. By Turing's original framing — if it is indistinguishable from a human, call it intelligent — ELIZA should count.
But ELIZA was not intelligent. It was not even close. It was a handful of pattern-matching rules applied to text. The "intelligence" existed entirely in the mind of the person using it, not in the program itself. Humans are extraordinary at finding meaning and intention in patterns — we see faces in clouds, hear emotion in minor chords, sense personality in a text message's punctuation. ELIZA exploited that tendency without doing anything to deserve it.
This is the crack in the Turing Test's foundation: it measures perceived intelligence, not actual intelligence. And perceived intelligence turns out to be much easier to produce than the real thing. A system does not need to understand language to produce language that sounds understanding. It does not need a model of the world to generate text that seems world-aware. It needs, in a sense, to understand us — our patterns, our expectations, our tendency to fill in gaps — well enough to exploit those tendencies.
That is why the field has largely moved away from the Turing Test as a meaningful benchmark. Passing the test tells us that a system can produce human-like output. It does not tell us what is happening inside the system, how it got there, or whether we should trust it with things that matter.
Strong AI and Weak AI
The ELIZA story also illuminates one of the fundamental debates in AI: the distinction between weak AI and strong AI.
Weak AI is the claim that machines can be programmed to exhibit intelligent behavior — to perform tasks that, if a human performed them, we would call intelligent. This claim is broadly accepted today and is demonstrated by every working AI system. A spam filter exhibits intelligent behavior. A chess program exhibits intelligent behavior. A medical imaging system exhibits intelligent behavior. No one seriously disputes this.
Strong AI is a much bolder claim: that machines can be programmed to actually possess intelligence — genuine consciousness, understanding, and inner experience, not just the outward appearance of it. This claim is hotly disputed and remains an open philosophical question.
The argument for strong AI goes something like this: the human mind is, after all, a physical system. It is built from neurons that individually do not think or feel, yet in combination they produce consciousness. Why should the same phenomenon be impossible in silicon? If the substrate does not matter — if what matters is the pattern of information processing — then a sufficiently complex machine could, in principle, be conscious.
The argument against strong AI goes like this: there is something irreducibly different about biological intelligence. A machine can simulate the outputs of intelligence without having the inner experience that produces them. No matter how convincingly ELIZA responded, it felt nothing. It knew nothing. It was not there in any meaningful sense. The same may be true of any machine, no matter how sophisticated.
We will not resolve this debate — philosophers and scientists have not resolved it in seventy years. But it is worth holding as an open question as we study the specific techniques in Week 6 and Week 7. Every AI system we examine is clearly weak AI. Whether any of them edge toward strong AI is a question that requires both technical understanding and philosophical humility.
The Question Is Older Than the Headlines
It is tempting to think of the current moment in AI as unprecedented. In some ways it is — the capabilities of modern large language models would have been unimaginable twenty years ago. But the core questions are not new.
In 1966, people interacted with ELIZA and believed it understood them. In 2024, people interact with ChatGPT and believe it understands them. The programs are incomparably more sophisticated. The human tendency being exploited is exactly the same. And the ethical questions Weizenbaum raised — about what it means to replace genuine human connection with a simulation of it, about who benefits from that replacement and who does not, about what we lose when we stop asking whether the system actually understands — are at least as urgent now as they were then.
Understanding the ELIZA story does not make you cynical about AI. It makes you precise. It gives you language for a distinction that matters: the difference between a system that produces intelligent-seeming output and a system that actually understands. That distinction is not always clear. But knowing it exists, and knowing it matters, changes the questions you ask — as a user, as a citizen, and as a teacher whose students are growing up with these systems around them all the time.
What Comes Next
With Topic 6a behind you, you have the conceptual vocabulary for the rest of Week 6: what an agent is, how agents make decisions, what kinds of knowledge they use, and the deep question of what it means for a machine to "understand" anything at all.
In Topic 6b, we look at one of the core techniques AI uses for goal-directed behavior: reasoning through problems by building and searching through a space of possible states. This is the foundation of everything from chess programs to route planners to puzzle-solving robots — and it is one of the clearest windows into how AI "thinks" when it is doing something more than reacting.