ELIZA is a conversation program written by Joseph Weizenbaum at MIT and published in 1966: the first famous chatbot, although, as Shrager (2024) argues, it was conceived not as a product but as a platform for studying communication between people and machines. It ran on the Project MAC time-sharing system, was written in MAD-SLIP for an IBM 7094, and took its name from the Eliza of Pygmalion: like her, its language could be gradually refined by a "teacher" (Weizenbaum, 1966, p. 36).
How it works
Nothing in ELIZA learns or reasons. The program scans the sentence for keywords, each with a precedence rank; if it finds one, it applies an associated decomposition rule and then a reassembly rule; if not, it produces a content-free remark or retrieves an earlier transformation (Weizenbaum, 1966, p. 37). Weizenbaum explains it with a foreigner who knows little English: he hears "I am very unhappy these days," understands only "I am," and, wishing to seem interested, replies "How long have you been very unhappy these days?" He applied a template and a "reassembly kit," whatever the rest happened to mean (1966, p. 37).
In formal notation, the template (0 YOU 0 ME), where 0 matches any sequence, decomposes "It seems that you hate me" into four parts, and the rule (WHAT MAKES YOU THINK I 3 YOU) reinserts the third (1966, pp. 37–38). These are production rules over strings, in the tradition of symbolic AI. A wink for linguists: a footnote explains that "transformation" is used in a generic sense, not in Harris's and Chomsky's (1966, p. 37).
The decisive point is that the script is data, not program: change the script and you change the persona, even the language. There were already scripts in Welsh and German (1966, p. 37), and the largest had about 50 keywords (1966, p. 38).
DOCTOR, the Rogerian therapist
The famous script, DOCTOR, imitates a Rogerian psychotherapist. This was no whim: the psychiatric interview is one of the few dialogues in which one participant can adopt "the pose of knowing almost nothing of the real world" (1966, p. 42). If someone says they went for a boat ride and the therapist replies "Tell me about boats," nobody concludes that he does not know what a boat is: we assume he has a purpose. That assumption is supplied by the speaker, who thereby dresses the program's replies in plausibility. Some subjects were very hard to convince that ELIZA was not human, which Weizenbaum calls a striking form of Turing's test (1966, p. 42) (see The Imitation Game).
Limits
Weizenbaum was its most lucid critic. ELIZA throws away nearly everything it is told, except what its small MEMORY keeps, and one of its main objectives had been to conceal its lack of understanding (1966, p. 43). The paper's conclusion sounds contemporary: ELIZA shows how easy it is to create and maintain the illusion of understanding, and with it of judgment deserving of credibility; "A certain danger lurks there" (1966, pp. 42–43). Years later, in Computer Power and Human Reason (Weizenbaum, 1976), he became a critic of handing judgment over to computers.
ELIZA and today's chatbots
An LLM-based assistant has no written script: its continuations are learned from an enormous corpus. But it keeps something of ELIZA's design. Shanahan (2024) points out that a bare-bones LLM is not a conversational agent: it must be embedded in a system that manages turn-taking and steered with a dialogue prompt (2024, p. 72). What changed is how much work the model does; what did not change is how much work the user does.