There is a moment, near the end of the ELIZA program, where the machine simply says "HMMM". Heise's "Zahlen, bitte!" column follows the source listing to line 404, close to the end of the code, and finds exactly that: a hand-written branch producing a thoughtful little noise.
It is a fitting artefact for the program that started the whole business. ELIZA was written by Joseph Weizenbaum at MIT in the mid-1960s, in a language called MAD-SLIP, and it ran on an IBM 7094 mainframe under the CTSS time-sharing system. Its most famous persona, DOCTOR, mimicked a client-centred psychotherapist: it picked up keywords in what a user typed, rearranged them with a handful of rules, and asked the question back. "I am unhappy." — "Do you think coming here will help you not to be unhappy?"
There was no comprehension anywhere in the program, and no attempt at it. Yet people who typed into ELIZA behaved as if there were — a phenomenon Weizenbaum named after himself the "ELIZA effect" and came to regard as an uncomfortable result of his own experiment. He would spend the rest of his career warning that some tasks should not be delegated to machines, regardless of what machines could do.
The rediscovery of the original ELIZA source code in an MIT archive a few years ago made it possible to read the program the way its creator wrote it, and the details are charmingly human. The dialogue tricks that convinced early users of a mind at work are, in the source, exactly what the "HMMM" suggests: a programmer's guess at how a listener sounds when it is listening. Modern language models produce soothing filler by prediction at enormous scale; ELIZA produced it by hand, on one line, and got the same effect.
