ELIZA: the chatbot that was never meant to be one (1966)

In 1966, Joseph Weizenbaum, a computer scientist at MIT, published a program that could hold what felt like a conversation. He called it ELIZA, after the character in Pygmalion who learns to speak properly. The program was simple, deliberately so. Its creator intended it as a demonstration of how shallow language processing really is. Instead, it became one of the most misunderstood inventions in the history of artificial intelligence.

How ELIZA worked

ELIZA was not one program but a framework. It ran interchangeable scripts, sets of pattern-matching and substitution rules, that dictated how it responded. The program itself, written in MAD-SLIP (later reimplemented in Lisp), scanned user input for keywords, ranked them by priority, broke the sentence apart using a decomposition rule, and reassembled a response.

The most famous script was DOCTOR, which simulated a Rogerian psychotherapist. Rogerian therapy was a brilliant choice: the therapist’s role is not to diagnose or advise but to reflect the patient’s words back to them. “You say you feel sad about your mother?”, that is both valid therapy and trivially easy to generate from a keyword match on “mother” and a grammatical reflection.

A typical exchange:

Human: My mother makes me angry. ELIZA: Tell me more about your family.

No understanding. No memory of previous turns beyond simple state. Just a keyword (“mother” → family), a decomposition pattern, and a reassembly template.

The ELIZA effect

What startled Weizenbaum was not that ELIZA worked, he knew exactly how shallow it was. What startled him was how people reacted.

His secretary, who knew the program was a set of rules, asked him to leave the room so she could talk to it privately. Students at MIT spent hours typing their problems into a terminal. Some users genuinely believed they were communicating with a being that understood them.

This phenomenon, attributing intelligence, empathy, or understanding to a system that has none, became known as the ELIZA effect. It is arguably Weizenbaum’s most lasting contribution, more important than the program itself.

Weizenbaum’s regret

Weizenbaum had designed ELIZA to show that conversation could be faked with trivial rules, hoping this would make people more skeptical of machines. The opposite happened. Psychiatrists proposed using ELIZA-style programs as actual therapeutic tools. Researchers cited it as progress toward machine understanding.

Horrified, Weizenbaum spent much of his later career warning against over-trusting computers. His 1976 book Computer Power and Human Reason argued that some tasks, particularly those involving empathy and moral judgment, should never be delegated to machines, regardless of technical capability.

Why it matters

ELIZA contained no neural network, no learning, no statistical model. It was pure symbolic AI manipulation: keywords, patterns, templates. And yet it convinced people.

Sixty years later, chatbots and large language models are incomparably more capable, but the ELIZA effect has not gone away. Users still project understanding onto systems that predict the next token. Weizenbaum’s uncomfortable question persists: when a machine sounds like it understands, whose understanding are we really measuring, the machine’s, or our own?