Curriculum · Unit 4 · Machines that talk

Why do people believe a machine understands them?

In 1966 a program made of a few dozen pattern rules got people to tell it their troubles. In 2025 a large language model was picked as the human in Turing's own three-way test more often than the actual humans were. The machinery changed completely. The people on the other side of the screen, it turns out, did not.

Nib says: First, talk to my pocket ELIZA. Type anything, or tap a Try sentence, then open Under the hood to see the exact rule it used. No understanding anywhere, I promise. Then press Next on the second machine to walk through the 2025 Turing test one step at a time.

Try itTalk to a mini ELIZA

Mini-ELIZA · 14 rules and a fallback
Try:

A teaching copy written for this page in ELIZA's style, not Weizenbaum's actual DOCTOR script. It swaps pronouns after matching, which is simpler than the original.

Step through itThe 2025 Turing test

    The idea in plain words

    ELIZA didn't understand a single word, and it didn't need to. Joseph Weizenbaum wrote it at MIT and described it in 1966. It worked from a script, and the famous one, called DOCTOR, played a Rogerian psychotherapist: the kind of counselor who mostly reflects your own words back and asks you to say more. That was a clever choice. A therapist who answers with questions never has to know anything about the world.

    Under the hood ELIZA did three things. It scanned your sentence for a keyword, picking the most important one if there were several. It used a pattern to cut the sentence into pieces. Then it dropped one piece into a ready-made reply, after flipping "I" and "my" to "you" and "your". If no keyword matched, it fell back on something like "Please go on." The toy above does exactly that with fourteen rules.

    People confided in it anyway. Weizenbaum later wrote that his own secretary, who had watched him build it, asked him to leave the room so she could talk to it in private. That troubled him deeply, and he spent much of the rest of his career warning that people too readily hand their trust to machines. The habit of reading understanding into a machine's words is now called the ELIZA effect. In 2025 a team even brought the original back to life: they restored the code from a printout found in Weizenbaum's papers at MIT and ran it on an emulation of the 1960s computer it was written for.

    A modern chatbot has no keyword list at all. It learned, during pretraining on enormous amounts of text, to predict the next token, and it writes every reply one token at a time. That makes it vastly more capable than ELIZA: it can talk about almost anything. It can also be fluent and wrong at the same time, which is called a hallucination.

    In 2025 Cameron Jones and Benjamin Bergen ran Turing's original three-party game. When GPT-4.5 was told to play a particular persona, people picked it as the human 73% of the time, more often than the real person in the other chat. That's a real milestone, and also a warning. It measures how convincing a machine is in a five-minute chat, not whether anyone is home. The ELIZA effect didn't go away. It got a much better machine to work with.

    Two ways to make a reply

    1966 · ELIZA

    Hand-written rules

    Find a keywordCut the sentence with a patternFlip I → you, my → yourSlot the piece into a canned reply

    Every possible reply was written by a person in advance. It can't say anything outside its script, so it rarely says anything false. It just doesn't say much.

    2020s · Large language model

    Learned prediction

    Turn the chat into tokensRun them through billions of learned weightsScore every possible next tokenPick one, add it, repeat

    Nobody writes its replies. It can talk about almost anything, in any style, including a persona. It can also be confidently wrong.

    What they share: both produce text that sounds like a person listening. The feeling of being understood is made on our side of the screen.

    Three things to keep

    Check yourself

    Why was a Rogerian therapist such a smart role for ELIZA?

    That style of therapist mostly reflects your words back and invites you to keep talking. ELIZA could sound attentive without knowing anything about the world, and when it said something odd, people assumed it was a deliberate therapist's question.

    You type "I am tired of my boss" into an ELIZA-style program. What happens, step by step?

    It spots the keyword "I am", takes the piece after it ("tired of my boss"), flips "my" to "your", and slots it into a reply such as "What is it like to be tired of your boss?" Try it in the toy above and open Under the hood.

    GPT-4.5 was judged to be the human 73% of the time. Does that show it understands people?

    No. It shows that in five-minute text chats, most interrogators couldn't tell it apart from a person, and often preferred it. Whether a machine understands is a different question that a short chat can't settle. ELIZA showed how easily we read understanding into fluent replies.

    Go deeper