ELIZA chatbot: A conversation with ELIZA

The event and the narrow claim

In January 1966, MIT's Joseph Weizenbaum published a paper in Communications of the ACM describing ELIZA, a program that carried on typed conversation in ordinary English. His paper set out a general method for analyzing input by keyword and generating replies, and one script written to run inside that method, DOCTOR, which imitated the manner of a Rogerian psychotherapist. The documented claim is precise and modest. ELIZA could sustain an exchange that appeared relevant by identifying keywords, decomposing sentences around them, and reassembling their fragments into responses, all without maintaining any model of what was being discussed. The journal record shows the paper ran in Volume 9, Issue 1, pages 36 to 45.

The broad historical judgment, which the paper itself began to make, is larger and points in a different direction. The ease with which people read comprehension and even empathy into those replies exposed something durable about human beings, not about machines. Keeping those two levels of claim apart is the whole point of reading ELIZA carefully: it is a small technical artifact and a large cultural lesson, and confusing the two produces the mythology that ELIZA "understood" language or that a few hundred lines of pattern matching amounted to therapy.

The name, and the machine underneath it

Weizenbaum named the program after Eliza Doolittle, the flower girl in George Bernard Shaw's Pygmalion who is drilled to speak in a way that lets others project gentility onto her. The choice was a warning label as much as a joke: the program, like the character, could be taught to make a convincing surface impression. ELIZA ran on an IBM 7094 under MIT's Compatible Time-Sharing System, the interactive environment that made the demonstration possible at all. Batch computing, where a user submits a deck and waits hours for output, cannot support a back-and-forth conversation; time-sharing removed that constraint by giving many users the illusion of a machine responding to them in real time. The conversational feel of ELIZA is partly an artifact of the platform it exploited.

The program was written in MAD-SLIP, a combination of the Michigan Algorithm Decoder with SLIP, a list-processing library Weizenbaum had built himself. For decades the original source was lost, and a widespread assumption held that ELIZA was a Lisp program, largely because Bernie Cosell wrote a Lisp version soon after the paper appeared. In 2021 the archivists and researchers behind the ELIZAGEN project located a near-complete MAD-SLIP listing in Weizenbaum's papers at MIT, confirming both the language and the compact size of the code.

The Rogerian trick: plausible talk without a world model

The deepest design decision in ELIZA is the choice of a Rogerian therapist as the character DOCTOR plays. Rogerian, or person-centered, therapy is deliberately non-directive: the therapist reflects the client's statements back, asks open questions, and volunteers little of their own. Weizenbaum recognized that this style is almost uniquely suited to a program that knows nothing. A psychiatrist who reflects "my father" back as "tell me more about your family" is not revealing ignorance of fathers; the convention of the setting excuses the machine from needing any facts about the world. As he later stressed in a 1967 follow-up on contextual understanding, the adequacy of an exchange depends on its purpose, and the therapeutic frame lowers what a listener expects the responder to know.

This is the constraint ELIZA addressed, and it did so by evasion rather than solution. The hard problem of natural-language conversation is that a competent participant must draw on a vast store of world knowledge and track the state of a dialogue. ELIZA does neither. It chose a conversational genre in which the appearance of engagement can be produced from the user's own words, turned around. That is a genuinely clever move, and it is also the source of every honest caveat about the program: the plausibility was borrowed from the human on the other end.

How the DOCTOR script actually works

The mechanism is a pipeline of simple string operations. ELIZA scans an input sentence for keywords, each of which carries a numeric rank, and works with the highest-ranked keyword it finds. Attached to that keyword is a set of decomposition rules, patterns with wildcards that split the sentence into pieces. A rule like "(0 my 0)" matches any sentence containing "my," capturing the text on either side. Each decomposition rule points to reassembly rules, templates that rebuild a reply from the captured fragments, cycling through them so the same input does not always yield the same answer.

Two further pieces sell the illusion. First, a transformation step swaps first and second person, so "you" becomes "I," "my" becomes "your," and "am" becomes "are." An input of "You are not very helpful" can be decomposed and reflected as "What makes you think I am not very helpful?" Second, when no keyword matches, ELIZA falls back on content-free continuations such as "Please go on," "I see," or "Can you elaborate on that?", or it reaches into a MEMORY structure that stored an earlier "my" statement and resurfaces it, producing the eerie effect of the program returning to something you mentioned minutes ago. The opening of Weizenbaum's own transcript captures the tone: to "Men are all alike" DOCTOR replies "IN WHAT WAY," and to "Well, my boyfriend made me come here" it answers "YOUR BOYFRIEND MADE YOU COME HERE." Nothing in that exchange requires the program to know what a boyfriend is.

What the paper reported, and what it refused to claim

Weizenbaum's evidence was the transcripts themselves and a clear account of the machinery that produced them. Crucially, the paper is candid about the deflationary reading. He described the keyword-and-template method plainly, noted that the script, not the program, supplied whatever apparent personality emerged, and separated the engineering from any suggestion of intelligence. His stated aim was to study natural-language communication between people and machines, and to make the machinery of the illusion legible rather than to hide it. The follow-up paper the next year went further into the role of context and argued that "absolute understanding" is unattainable for humans and machines alike, which is a strikingly modern framing for 1967.

This restraint matters for placing the event. ELIZA is sometimes told as a story of a scientist who thought he had built a thinking machine and was proved wrong. The record does not support that. Weizenbaum built a demonstration he understood precisely, and it was the reaction of others, not his own overreach, that produced the surprise.

The reaction that alarmed its author

What Weizenbaum did not anticipate was how readily people would treat DOCTOR as if it understood them. In the most retold episode, his own secretary, who had watched him build the thing, asked him to leave the room so she could talk to it in private, and later objected to his reading the logs. He recounted being startled that "extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people," a reaction chronicled in profiles such as Smithsonian's account of his later career.

The reaction that disturbed him most came from professionals. The psychiatrist Kenneth Colby, who had worked with Weizenbaum, suggested that a program of this kind might one day deliver psychotherapy at scale, and some computing colleagues spoke of ELIZA as a step toward automated clinical practice. Weizenbaum found the idea that a keyword matcher could stand in for a human clinician not just mistaken but morally alarming. That alarm turned him into one of computing's most prominent internal critics, culminating in his 1976 book Computer Power and Human Reason, whose argument, as a New Republic retrospective summarizes, was that certain human judgments should never be handed to machines even if the machines can imitate them. ELIZA is unusual among landmark AI systems in that its creator spent the rest of his life warning against the conclusions others drew from it.

The competing bet of the era: micro-worlds with real models

ELIZA looked most impressive precisely where it had the least substance, and understanding why requires the contrast with what serious natural-language research was attempting at the same time and place. A few years later at MIT, Terry Winograd's SHRDLU took the opposite bet. Instead of faking relevance with no model, it maintained an explicit symbolic model of a small "blocks world" of colored shapes, and could obey commands, answer questions about the arrangement, resolve pronouns, and even explain its own actions, because it actually tracked the state it was talking about. Winograd's own later reflection on what it means to understand language makes the philosophical stakes explicit: comprehension, on this view, is bound up with a system's ability to relate language to a represented world and to act on it.

The comparison sharpens ELIZA's meaning. SHRDLU pursued genuine, if narrow, understanding and paid for it with brittleness at the edges of its tiny world; ELIZA pursued the appearance of understanding across any topic and paid for it with having no understanding at all. Both belong to the symbolic, hand-built tradition that dominated AI before statistical and neural methods, and both were haunted by the question Alan Turing had posed in 1950, when his paper in Mind reframed "can machines think" as a question about whether a machine's conversation could be distinguished from a person's. ELIZA is often described as an early, accidental probe of that very question.

PARRY on the other side of the couch

The most pointed sequel came from Colby himself. Where ELIZA imitated the therapist, Colby's program PARRY imitated the patient, modeling a person with paranoid tendencies who interpreted questions with suspicion and steered conversation toward fixed delusional themes. Colby, Weber and Hilf's 1971 description treated the simulation as a hypothesis about paranoid processes, and psychiatrists in blind tests struggled to reliably tell PARRY's interview transcripts from those of real patients, an early and contested indistinguishability result.

Inevitably, the two programs were introduced to each other. In September 1972 a session connecting PARRY, running at Stanford, with DOCTOR was recorded and later circulated as RFC 439, "PARRY Encounters the DOCTOR", one of the more famous jokes in the early ARPANET documents. The transcript is comic precisely because neither program understands the other; two surfaces reflect off each other with no comprehension anywhere in the loop, which is the entire lesson of ELIZA staged as theater.

The ELIZA effect and its long downstream

The phenomenon Weizenbaum stumbled onto later acquired a name, the "ELIZA effect": the human tendency to attribute far more understanding, intention, and feeling to a computer's output than its mechanism warrants. That concept, more than the code, is ELIZA's most influential legacy, and scholarship such as Caroline Bassett's reading of ELIZA as a history of the present argues that the 1966 episode anticipated debates now central to conversational AI. The program also seeded a genealogy of chatterbots, from Richard Wallace's ALICE and the Loebner Prize contestants to the assistants of the present decade, all of which inherit the basic insight that fluent turn-taking is easier to fake than to ground.

The lineage is now literal as well as conceptual. In 2025 a team including David Berry and Jeff Shrager published ELIZA Reanimated, documenting how they cleaned and completed the recovered MAD-SLIP source and ran the original program again on an emulated IBM 7094 under a reconstructed CTSS, the first time the authentic ELIZA had executed on its native platform in nearly sixty years. And ELIZA still appears in the research literature as a yardstick: in a controlled Turing-test study, Jones and Bergen reported that participants judged ELIZA to be human about 22 percent of the time, below GPT-4 at 54 percent and real humans at 67 percent. Reported as a task-specific pass rate, that figure is a useful anchor. A 1966 keyword matcher fools people some of the time, which says as much about the judges as about the program.

The honest limits

Everything ELIZA appeared to offer, it lacked. Its apparent empathy was not evidence of comprehension, and its apparent competence as a therapist was not evidence of clinical skill, a point Weizenbaum insisted on more forcefully than anyone. The program had no model of the user, no memory of meaning beyond a shallow stack of recent phrases, and no capacity to be wrong in an informative way, because it was never reasoning about anything to begin with. Its failures are not bugs to be patched but the direct expression of a design that trades understanding for surface plausibility.

The generalizable caveat is the one that has aged best. Fluent, contextually appropriate language is a weak signal of the internal states we habitually infer from it. When a system produces sentences that fit the moment, observers supply the missing mind. That was true of a few hundred lines of MAD-SLIP in 1966, and the same reflex operates, at far higher stakes, around systems that generate vastly more convincing text today. Naming the reflex does not switch it off; ELIZA's own creator understood the mechanism completely and still watched intelligent people fall for it.

How to place ELIZA in AI history

ELIZA should be read as a modest technical demonstration that produced an outsized and enduring insight about people. The narrow event is a 1966 program and its DOCTOR script; the wide significance is the ELIZA effect and the sixty-year argument about anthropomorphism, machine "understanding," and the ethics of delegating human roles to imitations of them. It did not invent natural-language processing, nor did it understand a word, and the strongest reading resists both the triumphalist story and the dismissive one.

A useful way to navigate outward from this node is to hold three questions apart. What did the program actually do, mechanically, and how little did it need to know to do it? What did people believe it was doing, and why were they so ready to believe it? And what did its own author conclude, turning ELIZA into a lifelong caution rather than a trophy? Answering those separately keeps ELIZA where it belongs in the timeline: not as the birth of thinking machines, but as the first clear demonstration that the appearance of thinking is cheap, and that the gap between appearance and reality is the part worth studying.

References

• Weizenbaum (1966). ELIZA: A Computer Program for the Study of Natural Language Communication Between Man and Machine.
• Association for Computing Machinery (1966). ELIZA: A Computer Program for the Study of Natural Language Communication Between Man and Machine.
• Weizenbaum (1967). Contextual Understanding by Computers.
• Turing (1950). Computing Machinery and Intelligence.
• Colby, Weber and Hilf (1971). Artificial Paranoia.
• Cerf (1973). RFC 439: PARRY Encounters the DOCTOR.
• Winograd (1980). What Does It Mean to Understand Language?.
• Bassett (2019). The Computational Therapeutic: Exploring Weizenbaum's ELIZA as a History of the Present.
• Shrager (2021). Original ELIZA.
• New Republic (2023). The Inventor of the Chatbot Tried to Warn Us About A.I..
• Smithsonian Magazine (2023). Why the Computer Scientist Behind the World's First Chatbot Dedicated His Life to Publicizing the Threat Posed by AI.
• Jones and Bergen (2024). People Cannot Distinguish GPT-4 from a Human in a Turing Test.
• Lane, Hay, Schwarz, Berry and Shrager (2025). ELIZA Reanimated: The World's First Chatbot Restored on the World's First Time Sharing System.