ELIZA: The 1960s MIT Chatbot That Horrified Its Own Creator

Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode!

In the mid-1960s at MIT, Joseph Weizenbaum walked into his lab and found his secretary typing at a teletype terminal. She asked him to leave the room. She wanted privacy for a personal conversation with ELIZA, the program he had written between 1964 and 1967 and named after Eliza Doolittle of Pygmalion. Weizenbaum had built it to show that conversation between people and machines was a superficial trick. Instead, users confided in it, and psychiatrists proposed running it on many terminals to treat hundreds of patients at once.

This episode explains exactly how the illusion worked. The famous DOCTOR script played a Rogerian psychotherapist, a persona that excuses knowing nothing about the world. The program ranked keywords, chopped sentences apart, swapped pronouns, and pasted the fragments into stock templates. Weizenbaum was appalled by what followed and spent years arguing that calculation is not judgment, a warning that reads differently in the age of large language models.

• Weizenbaum wrote ELIZA in MAD-SLIP, a list processing language he developed himself, and the program could not learn anything unless a programmer manually wrote new script rules.
• Type You are very helpful and the DOCTOR script flips the pronouns and answers What makes you think I am very helpful, with a memory queue for lines like Earlier you mentioned your father.
• The ELIZA effect names the bias that makes people credit a program with understanding and feeling even when they know it is only matching text patterns.
• At a 1972 computer communications conference, ELIZA was connected to PARRY, a program built by a Stanford psychiatrist to simulate a paranoid patient, and the two scripts talked past each other.
• The original source code was lost for decades until Jeff Shrager and an MIT archivist found it in 2021 in a folder labeled Computer Conversations, and in 2024 a team led by Rupert Lane ran it on an emulated IBM 7094.