ELIZA, 1966: the chatbot that answered every feeling with a question

The sample conversation is printed right there on the first page. Someone types "Men are all alike." The machine, in capital letters, answers "IN WHAT WAY". A few lines later the person mentions a boyfriend, and the program hands the word back as a question. It reads like a patient talking to a very calm therapist. It was a program running on a time-sharing system at MIT.

That page belongs to a paper Joseph Weizenbaum published in Communications of the ACM, Volume 9, Number 1, in January 1966. The program was ELIZA, and almost every chatbot argument we still have started with it.

How the trick worked

Weizenbaum was open about the machinery, and it's simpler than the conversation makes it look. ELIZA ran on MIT's MAC time-sharing system and was written in a language called MAD-SLIP for the IBM 7094. When you typed a sentence, it looked for a keyword. Keywords had ranks, so a more important one could win over a weaker one found earlier in the line.

Each keyword came with a decomposition rule, which chopped your sentence into parts, and a reassembly rule, which put some of those parts into a stock reply. The paper's own example: a sentence of the form "I am BLAH" can become "How long have you been BLAH", without the program knowing what BLAH means. If nothing matched, ELIZA fell back on a content-free remark or, under certain conditions, reused an earlier transformation.

The detail I find most interesting is one line about scripts. Weizenbaum called the set of keywords and rules a SCRIPT and pointed out that the script is data, not part of the program. Swap the script and you get a different character. Scripts existed, he noted, in Welsh and German as well as English. The famous one, DOCTOR, imitated a Rogerian psychotherapist, a style built on turning statements back into open questions. That's why the therapy setting worked so well. A therapist who mostly reflects your words back is the one conversation partner a keyword matcher can play convincingly.

The part that bothered its maker

People believed it anyway. Wikipedia recounts that Weizenbaum's secretary reportedly asked him to leave the room so she could talk to ELIZA in private. He later wrote that very short exposure to a simple program could induce powerful delusional thinking, and his 1976 book, Computer Power and Human Reason, argued against handing human judgment over to computers. The tendency he worried about now has his program's name on it: the ELIZA effect.

My honest take is that the ELIZA effect is the most useful idea from 1966, more useful than the code. Today's models are not keyword matchers. They are far more capable. But fluent replies still make people assume understanding, and that assumption is where most bad AI decisions start.