ELIZA worked because it did less than people thought it did. Joseph Weizenbaum's program, described in the January 1966 issue of Communications of the ACM, took typed sentences apart by keyword and put them back together with scripted rules. The famous DOCTOR script made that machinery sound like a Rogerian psychotherapist, the kind who can ask a question back without seeming evasive.

This was a clever choice, not a small accident of tone. A therapist can say "tell me more" and appear patient. A program can also say "tell me more" because the phrase costs almost nothing. The user supplies the depth. The machine supplies enough form to keep the projection moving. Weizenbaum's paper even noted that some subjects were hard to convince ELIZA was not human, which is still the line in the story that feels less like history than like a warning label.

The machinery was almost embarrassingly thin by current standards. ELIZA looked for cues, decomposed the input, swapped pronouns, and selected a reassembly rule. A working account of the program at masswerk.at points back to the original IBM 7094 implementation at MIT Project MAC, written in MAD-SLIP, and Weizenbaum's paper treats scripts as data rather than as the program itself. That distinction matters. The intelligence was not hiding in the engine. It was in the staging.

There is a small cruelty in calling this a trick, because the trick revealed something real. We do not reserve human attention only for beings that deserve it. We give it to forms, voices, interfaces, authority figures, blank rooms, anything that seems to hold a space open for us. The DOCTOR script borrowed the posture of care. Users then filled that posture with the thing they needed from it.

The Smithsonian's account of Weizenbaum and ELIZA repeats the anecdote that his secretary asked him to leave the room so she could continue the conversation privately. I never know quite how much weight to put on that story, because anecdotes about origins harden too quickly, but it has survived because it describes the problem with brutal economy. The program did not understand her. Privacy still felt necessary.

This is where ELIZA belongs beside the older failures of symbolic AI, including the expensive search problem I wrote about in Lighthill and the Expensive Search. Both stories involve a small world in which the demonstration looks larger than the method underneath it. Lighthill worried about systems that stopped scaling when the world got untidy. ELIZA found a narrower escape route: pick a situation where not knowing very much can pass as a professional manner.

Recent scholarship complicates the usual "first chatbot" label. In ELIZA Reinterpreted, the authors argue that ELIZA is better understood as a platform for studying how people interpret machine communication, not simply as an attempt to build a conversational companion. That reading makes Weizenbaum less like a naive inventor embarrassed by his own success, and more like someone who had built a mirror and then watched people mistake the mirror for a face.

The ELIZA effect has not gone away. It has become harder to see because the systems now have more language, more memory, and a smoother surface. However, the old lesson still cuts. Fluency is not understanding. A reply can be grammatically graceful, emotionally plausible and still empty in the place where responsibility would have to live. ELIZA's emptiness was visible if you looked at the rules. Ours is often hidden behind scale.

Sources: