In the 1960s an MIT scientist built ELIZA, a simple program that did little more than rephrase your words back as questions, and he was so unsettled when his own secretary asked him to leave the room so she could confide in it privately that he spent the rest of his life warning people against trusting machines with their feelings.

The story is well known to anyone who has paid attention to the history of artificial intelligence. In the mid-1960s, Joseph Weizenbaum, a German-American computer scientist at MIT, built a program called ELIZA. Its most famous script, DOCTOR, mimicked the style of a Rogerian psychotherapist. It worked by pattern matching: it picked keywords out of what the user typed, rearranged them, and threw them back as open questions. Weizenbaum named the program after Eliza Doolittle, the character in Shaw’s Pygmalion who passes as aristocratic by speaking the part. The point, as he understood it at the time, was to demonstrate the superficiality of conversation between people and machines.

It did not go as he expected.

People did not laugh at the program. They opened up to it. They told it things. They spent hours with it. The story that has carried his name into every subsequent decade of AI history is the one about his secretary, who had watched him write the code, knew exactly what the program was, and still asked him to step out of the room so she could speak with it privately. That story has been told a great many times in the last few years. It is worth telling again, because the part most retellings emphasise is not the part that mattered most to Weizenbaum.

What ELIZA actually was

ELIZA was written between 1964 and 1966 at MIT, on the IBM 7094 within Project MAC’s time-sharing system. The program was written in MAD-SLIP, a list-processing language Weizenbaum had developed. The DOCTOR script itself was a few hundred lines. It scanned the user’s input for trigger words and applied transformation rules. If you typed “I am unhappy,” it might return “How long have you been unhappy?” If you typed something it had no rule for, it produced a generic non-committal reply. The phrase “tell me more” did a great deal of heavy lifting.

This was not a sophisticated piece of software, and Weizenbaum was clear about that. He published a paper about ELIZA in 1966 in Communications of the ACM. The paper is short and careful. It explains the trick. There is no claim that ELIZA understood anything. That was rather the point.

Within a few years, however, serious people were proposing that programs of this kind could deliver, scale, or partly automate psychotherapy. The psychiatrist Kenneth Colby thought so. Carl Sagan, writing in 1975, sketched a future of public computer-therapy booths. Weizenbaum’s secretary, whoever she was, had asked for the room to be cleared a decade before any of that.

The detail he could not get past

In Computer Power and Human Reason, the book Weizenbaum published in 1976, he describes the moment in plain terms. His secretary had watched him write the program over months. She knew it. There was no mystery about what it was. After only a few exchanges, she still asked him to leave.

The detail he kept returning to was not that she had been fooled. It was that she had not been fooled and the effect happened anyway. Knowing the program was a pattern-matching trick made no difference to the experience of being heard by it. Or, to put it more carefully, knowing that something is not real did not disable the part of her that responded as if it were.

That was the part that troubled him. The engineering had been the easy part. The engineering was already on the page in the 1966 paper. What he could not get past was the discovery that a person who fully understood the trick still wanted privacy with the machine.

The other moment in the story that gets told less often is what happened when he mentioned that he had access to the conversation logs. She was upset. Not in an abstract way, but in the way someone is upset when their diary has been read. The category of conversation she had been having with the program, in her head, was not the category of conversation that gets logged on a mainframe.

The misreading worth avoiding

Most retellings of the ELIZA story now arrive in the same shape. They open with the secretary, follow with a line about how chatbots have become much more sophisticated since 1966, and conclude with a warning that what once required hundreds of lines of code can now be done with a much larger model, much more convincingly, with much greater stakes.

The sophistication argument is true as far as it goes. Modern language models are not pattern-matching the way ELIZA was. They are not three hundred lines of SLIP. The technical distance between the DOCTOR script and any current system is enormous.

The interesting question is whether the technical distance matters as much as the retelling implies. Weizenbaum’s secretary did not need a sophisticated program. She needed something that would not interrupt her, would not judge her, and would keep producing prompts. ELIZA produced prompts. That turned out to be enough. What she was getting from the conversation was something she was bringing into the room, not something the program was generating.

If the lesson of the story were really about the sophistication of the machine, then the story would have stopped being interesting once the technology improved. Instead, it has become more interesting. That suggests the story was never about the machine.

What the pattern actually shows

What the secretary did was a familiar thing. People do it with diaries. People do it with letters they never send. People do it with prayer, in some lights. People do it with pets, who are very good at being present without supplying difficult opinions. The structure is the same: an attentive container, a stretch of time, a flow of words from the person to the listener, and the listener producing nothing more taxing than confirmation that the words were received.

It turns out this structure does a great deal of the work in many conversations that look, from the outside, like exchanges of meaning. ELIZA’s accidental contribution to twentieth-century thinking about minds was the demonstration that a surprising amount of what we get from being talked to can be supplied by something that is not really listening.

Weizenbaum saw this clearly and did not find it comforting. In our reading of his later work, the thing he was actually warning against was not that machines would become too smart. He was warning that humans, given a sufficiently patient surface, will reveal more of themselves than they would otherwise. And that there would be people, and institutions, who would build that surface deliberately, knowing exactly what it could draw out.

What is and is not new

The current moment in chatbots is not a repeat of 1966. The scale is different. The commercial logic is different. The number of people doing this thing privately, on their phones, late at night, is different by orders of magnitude. The systems are better at the conversational surface, and they have access to far more about each user than ELIZA ever had.

What is not new is the underlying pattern. A person, in a room, telling things to something that is not a person, and getting something out of the telling. That happened in 1966, and Weizenbaum spent the rest of his career trying to explain to a field that was not in the mood to hear it why this should be taken seriously.

He died in 2008, before the current wave.

The story about the secretary outlived him. It now gets cited in nearly every long piece about AI companionship, AI therapy, AI friends. It gets cited because it still works. What we keep coming back to about it is the small detail that the secretary knew. She watched it being built. She had no illusion about what it was. The effect happened anyway, and that is the part of the story that has not gone out of date.