event · event/eliza
ELIZA
Also called the DOCTOR script
A record as of 2026-08-28. No longer actively maintained.
The first program that people mistook for a person was published as a warning about mistaking programs for people. Joseph Weizenbaum's paper in the January 1966 Communications of the ACM describes ELIZA, running on "the MAC time-sharing system at MIT," and opens with the conversation that made it famous: "Men are all alike. / IN WHAT WAY? / They're always bugging us about something or other. / CAN YOU THINK OF A SPECIFIC EXAMPLE?"
The mechanism is stated in two sentences: "Input sentences are analyzed on the basis of decomposition rules which are triggered by key words appearing in the input text. Responses are generated by reassembly rules associated with selected decomposition rules." There is no model of the world anywhere in it. Weizenbaum chose the therapy setting precisely because it excuses that absence: "the psychiatric interview is one of the few examples of categorized dyadic natural language communication in which one of the participating pair is free to assume the pose of knowing almost nothing of the real world." A script — the famous one makes ELIZA "respond roughly as would certain psychotherapists (Rogerians)," and came to be known as DOCTOR — supplies the keywords and the pose; the program supplies nothing else.
Weizenbaum named the program for Eliza Doolittle, and the name carries his whole thesis: "Like the Eliza of Pygmalion fame, it can be made to appear even more civilized, the relation of appearance to reality, however, remaining in the domain of the playwright." His stated aim in publishing was demystification: "once a particular program is unmasked, once its inner workings are explained in language sufficiently plain to induce understanding, its magic crumbles away." The paper is sixty years old and that sentence has not needed revision.
The program itself then vanished for most of those sixty years. What
circulated — in BASIC listings, in Emacs's M-x doctor — were
reimplementations of reimplementations. The original was written in
MAD-SLIP for MIT's CTSS, and its source survived only as paper: "We
discovered an original ELIZA printout in Prof. Weizenbaum's archives at
MIT, including an early version of the famous DOCTOR script," reports the
team — Rupert Lane, Anthony Hay, Arthur Schwarz, David M. Berry and Jeff
Shrager — whose ELIZA Reanimated, posted 12 January 2025, describes "the
reanimation of this original ELIZA on a restored CTSS, itself running on an
emulated IBM 7094. The entire stack is open source." The first chatbot now
runs again on an emulation of the machine it was written for.
The strangest entry in ELIZA's record is recent. Cameron Jones and Benjamin Bergen ran a public online Turing test, first posted on 31 October 2023; in the revised version of 20 April 2024, "the best-performing GPT-4 prompt passed in 49.7% of games, outperforming ELIZA (22%) and GPT-3.5 (20%), but falling short of the baseline set by human participants (66%)." Read that middle clause again: a keyword-matching script from 1966 was judged human slightly more often than GPT-3.5, the model behind the original ChatGPT. The first posting had put ELIZA ahead by a wider margin still — 27% against GPT-3.5's 14% — so the version you read changes the gap, not the order. Weizenbaum's playwright framing holds: the relation of appearance to reality stayed in the domain of the playwright, and the audience kept doing half the work.
Facts
- paper citation
- Weizenbaum, Communications of the ACM, volume 9, number 1, January 1966source, accessed 2026-08-28
- platform
- the MAC time-sharing system at MITsource, accessed 2026-08-28
- original source language
- MAD-SLIP, found as a printout in Weizenbaum's archives at MITsource, accessed 2026-08-28
- turing test success rate
- 22%, against 20% for GPT-3.5 (Jones and Bergen, revised version of 20 April 2024)source, accessed 2026-08-28
Timeline
- the ELIZA Reanimated team reports running the rediscovered original MAD-SLIP ELIZA on a restored CTSS on an emulated IBM 7094source
- the revised version restates the figures: ELIZA passes as human in 22% of games, ahead of GPT-3.5 at 20%source
- Jones and Bergen post the first version of their public Turing test, in which ELIZA passes as human in 27% of games, ahead of GPT-3.5 at 14%source
- Weizenbaum's ELIZA paper appears in the January 1966 Communications of the ACMsource