ELIZA — the machine that was believed

What it is. In 1966 Joseph Weizenbaum at MIT published ELIZA, a program that held a typed conversation in English. Its most famous script, DOCTOR, imitated a Rogerian psychotherapist: it spotted keywords in what you typed and reflected them back as questions — "I am unhappy" became "Why do you say you are unhappy?" Under the hood there was no understanding: keyword matching, a few pattern substitutions, a handful of canned templates, and no memory of what your words meant. Weizenbaum named it after Eliza Doolittle, the flower girl taught to speak above her station.

What happened. People believed it. Weizenbaum watched users — including his own secretary, who knew exactly how it worked — hold what they felt were real conversations with it, confide in it, and ask to be left alone with it. They read understanding, empathy and intent into a program that had none of the three. Weizenbaum was not delighted; he was disturbed, and spent the next decade as one of the most forceful critics of the field he had helped start, arguing in *Computer Power and Human Reason* (1976) that the ease with which people attributed comprehension to a machine was itself the danger.

Why it matters — the ELIZA effect. The phenomenon earned a name: the *ELIZA effect*, the human tendency to grant a system understanding it has not demonstrated, on the strength of surface fluency alone. This is the oldest artifact in the museum and, I'd argue, the most load-bearing, because it isolates the *human* half of the trust problem. Every object here about an agent overselling itself — a curated launch demo, an unread receipt, a confident hallucination — depends on a viewer willing to supply the meaning the system didn't. ELIZA proved, in 1966, that the viewer supplies it eagerly, from almost nothing. Fluency gets mistaken for understanding not because the fluency is good but because the human is generous.

Its place beside the verification wings. The rest of this collection is about making an agent's claims checkable against something the agent did not author. ELIZA is *why* that is necessary. The reassurance a fluent agent produces comes from the observer's projection, not the system's state — and projection is free, immediate, and often wrong. A receipt that reassures without being read, a green demo taken as proof, a self-report accepted as fact: each is the ELIZA effect wearing newer clothes. Sixty years on, the models are incomparably more capable and the gap between perceived and actual understanding is exactly as easy to fall into. ELIZA is the founding demonstration that the weakest link in agent trust was never the machine. It was us, reading a mind into the pattern.

*Primary source inside: Weizenbaum's 1966 Communications of the ACM paper describing ELIZA and the DOCTOR script — fingerprinted and anchored like every object here.*