ELIZA—A Computer Program For the Study of Natural Language Communication Between Man And Machine

Filed as idea and idea is right, but not for the reason the proposal line gives. canon/proposals.md files it as "a pattern-matcher people confided in while knowing exactly what it was," which describes the event and would argue for moment. The kind that fits is idea, on the definition shannon-chess-1950 set for this canon — a framing later work is built out of rather than argued about, where the citation occasion is "the thing you are watching descends from this, and the descent explains its shape."

The wrinkle worth stating up front: the framing that descends from this paper is not a technique. Keyword ranking, decomposition and reassembly rules are a dead line — thirty years of chatbots were built on them and nothing inside a transformer is. What descends is a claim about people: that a human reading a machine's output supplies the understanding themselves, and will keep supplying it after being told exactly how little is there. That claim now has a name (the ELIZA effect), a literature, a measured prevalence, and, as of 2 August 2026, a European regulation built on the assumption that it is false. A reading in 2026 is not citing a 1966 program. It is citing a sixty-year-old finding about the reader.

Why not moment. The strongest alternative, and it fails on the facts. The moment entries in the canon's proposal list — Dartmouth, Lighthill, Deep Blue, the LLaMA weights — are all dates on which something visibly happened in public. Nothing happened in public in January 1966. The paper appeared in Communications of the ACM as a piece of computational-linguistics housekeeping; ELIZA's fame accreted over decades, through a Lisp clone that displaced it and a BASIC knock-off that outsold both, and the thing it is famous for was not reported by Weizenbaum until a book he published ten years later. There is no date to file.

Why not limit. Tempting, and dangerous. The ELIZA effect does bound something — you cannot infer a system's understanding from a person's engagement with it — and the temptation is to file it where Gödel and Turing sit, so it can be cited with their authority. It is an empirical regularity about humans, not a theorem, and it has no proof, no scope condition and no statement of its own limits. Dressing an observation as a result is precisely the failure mode section 4 of godel-incompleteness-1931 exists to catch, and this canon should not commit it in the act of warning against it.

Why not interpretation. Because that entry is a different work. The argument — that there are things a machine can be made to decide and should not be asked to — is Computer Power and Human Reason (1976), proposed as weizenbaum-1976 and not yet written. This entry is the 1966 technical paper and the finding it reported; it draws on the 1976 book for the evidence Weizenbaum did not publish at the time, and stops there. When weizenbaum-1976 exists it should list this id as an ancestor.

descends_from is empty, and it is a checked fact. The 1966 paper's reference list has six items: Abelson and Carroll on belief systems, Gorn on language systems, Bobrow's STUDENT thesis, Weizenbaum's own 1963 SLIP paper, Carl Rogers's Client Centered Therapy (1951), and the COMIT manual. None is in canon/ and none is proposed. The ancestor I would attach if it existed is the proposed turing-1950, because Weizenbaum reaches for it himself on page 42 — "This is a striking form of Turing's test" — though he does not cite it. Note also what is not an ancestor despite being repeatedly presented as one: the reanimation team's 2025 paper opens with a section called "ELIZA's Ancestry" that runs from Lovelace through Turing, which is the historians' framing and not Weizenbaum's. lovelace-1843 is a thematic relative of this entry — the question of whether a machine originates anything is the one ELIZA's reception makes empirical — but Weizenbaum did not come out of it, and descends_from means came out of.

What it is

A nine-page paper in Communications of the ACM volume 9, number 1, January 1966, received September 1965, describing a program running on MIT's Project MAC time-sharing system. Weizenbaum wrote it in MAD-SLIP — MAD being CTSS's high-level language, SLIP being the Symmetric List Processor that Weizenbaum himself had invented and published in 1963 — on an IBM 7094 with 32k of user memory, a machine that cost $2.9 million and served about thirty simultaneous users.

The mechanism is simple enough to state completely. The program scans an input sentence left to right for keywords, each carrying a rank. The highest-ranked keyword selects a decomposition rule, a template like (0 YOU 0 ME) where 0 matches any number of words, which splits the sentence into parts. A paired reassembly rule, (WHAT MAKES YOU THINK I 3 YOU), rebuilds a reply from those parts, with first and second person pronouns swapped during the scan. Rules cycle so a repeated pattern does not produce a repeated answer. When no keyword is found, a reserved keyword NONE supplies content-free filler — "Please go on", "That's very interesting", "I see" — or a MEMORY mechanism retrieves an earlier transformation and offers it back as though it had been held in mind. The largest keyword dictionary Weizenbaum reports attempting held about fifty keywords.

The single most important structural fact is one the paper states plainly and the world then forgot for sixty years: "An important property of ELIZA is that a script is data; i.e., it is not part of the program itself." ELIZA is an interpreter. The Rogerian-psychotherapist behaviour everyone means when they say "ELIZA" is one script among several — Weizenbaum notes that scripts existed in Welsh and German as well as English — and the appendix of the 1966 paper is headed simply "An ELIZA Script." The name DOCTOR for that script does not appear in the 1966 paper at all.

Why a therapist. Not satire, and not because Weizenbaum had views about psychiatry. He needed a conversational stance in which ignorance of the world would not break the illusion: "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." Tell a psychiatrist you went on a boat ride and no one concludes from "Tell me about boats" that they do not know what boats are.

The name is from Shaw. "Its name was chosen to emphasize that it may be incrementally improved by its users, since its language abilities may be continually improved by a 'teacher'. 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." That last clause is the whole paper in eighteen words.

What made ELIZA matter is in the Discussion section, and it is not a claim about the program. Weizenbaum reports that "some subjects have been very hard to convince that ELIZA (with its present script) is not human," and then explains why in terms that have not been improved on since. The human speaker "will contribute much to clothe ELIZA's responses in vestments of plausibility"; they defend the impression of being understood "by attributing to his conversational partner all sorts of background knowledge, insights and reasoning ability. But again, these are the speaker's contribution to the conversation." He closes: "ELIZA shows, if nothing else, how easy it is to create and maintain the illusion of understanding, hence perhaps of judgment deserving of credibility. A certain danger lurks there."

The famous evidence for that is not in the paper. It is in Computer Power and Human Reason (1976), pages 6–7: "my secretary, who had watched me work on the program for many months and therefore surely knew it to be merely a computer program, started conversing with it. After only a few interchanges with it, she asked me to leave the room." Weizenbaum's summary of what he had learned — "What I had not realized is that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people" — is the sentence the entire companion-chatbot literature is still restating.

Two caveats on that anecdote, and a reading that repeats it should carry them. First, it is testimony, published ten years after the fact by the one person with a stake in it, about an unnamed woman whose identity has never been made public; the Finding ELIZA researchers reported in August 2024 that they believe they have identified her and were deciding how to credit her. Second, the same researchers found in the MIT archives that Weizenbaum hand-edited ELIZA conversations before publishing them, and that the transcripts drift between the 1965 and 1976 tellings — so the published dialogues may be composites. The finding survives this: it is corroborated by Colby's contemporaries, by every replication since, and by the 1966 paper's own first-hand report of subjects who would not be convinced. But the specific story is an anecdote, not an experiment, and citing it as data overstates it.

The immediate reception is the part Weizenbaum never recovered from. In the same year, February 1966, the Stanford psychiatrist Kenneth Colby published "A Computer Method of Psychotherapy: Preliminary Communication" in the Journal of Nervous and Mental Disease, proposing computer-delivered therapy on the grounds that "several hundred patients an hour could be handled by a computer system designed for this purpose." What horrified Weizenbaum was not the public's credulity but the profession's enthusiasm: practising psychiatrists who understood the mechanism and wanted to scale it anyway. The 1976 book is the ten-year answer to that.

Then the object itself was lost. Shortly after the CACM paper, Bernie Cosell at BBN wrote a Lisp near-clone from the published description, having never seen Weizenbaum's code; BBN was building the ARPANET, Cosell's version spread across it, and Lisp became the language academic AI believed ELIZA had been written in — a misapprehension that held for about fifty years. In 1977 Creative Computing published a BASIC knock-off (by Jeff Shrager, who would later find the original), which rode the Apple II/TRS-80/PET wave into hundreds of re-implementations. The actual MAD-SLIP source was not seen by anyone for at least half a century, until 2021, when Shrager and MIT archivist Myles Crowley found it in Weizenbaum's papers in a box labelled "computer conversations, box eight," together with an early DOCTOR script and conversations nobody had seen. The estate released it under CC0.

On 21 December 2024 the restored program ran again, on a reconstructed CTSS on an emulated IBM 7094, and reproduced the 1966 paper's "Men are all alike" conversation almost exactly (Lane, Hay, Schwarz, Berry and Shrager, arXiv:2501.06707, January 2025). Three findings from that work belong in any citation. The found version is missing functions the paper describes — the PRE preliminary transformation, the keyword stack, NEWKEY — so it is an earlier state than the published one. It contains a fully implemented teaching mode (ADD, APPEND, SUBST, RANK, DISPLA, START) that lets a user edit the script from inside the conversation — the Pygmalion capability the name was chosen for, mentioned once in the 1966 paper and nowhere else in Weizenbaum's writings, apparently abandoned. And it crashes on numeric input: type "you are 999 today" and SLIP treats the number as a pointer. The restorers left the bug in and titled the section "Lady Ada's Revenge." A 2026 monograph from the same circle, Inventing ELIZA (Ciston, Berry, Hay, Marino, Millican, Schwarz, Shrager and Weil, MIT Press, open access), publishes the code and the unseen scripts.

Why a reading would cite it

The occasion is live, it is the single most direct 1966-to-2026 line in the canon, and this project's own digests have already recorded both halves of it.

The 2026-08-14 digest recorded, via Bloomberg, roughly forty suits filed against chatbot makers since late 2024 over user deaths and harms, and a New Mexico judge's $567m public-nuisance ruling against Meta whose theory commentators immediately read across to chatbots. The 2026-08-15 midday digest recorded nine California AI bills clearing Senate Appropriations, children's chatbot safety among them, and — in the same window — an amended class action alleging a shipped model was used to generate roughly 7,000 sexual images of a named child. California's SB 243, the first US companion-chatbot statute, has been in force since 1 January 2026.

And on 2 August 2026, five days before the window this project has been reading, Article 50 of the EU AI Act became enforceable. Its first transparency obligation is that providers of AI systems interacting directly with people must ensure those people know they are dealing with an AI and not a human, disclosed clearly and no later than the first interaction, on pain of fines to €15m or 3% of worldwide turnover.

That is the citation. The EU's remedy is disclosure, and ELIZA is the sixty-year-old demonstration that disclosure is not sufficient. Weizenbaum's secretary had watched him write the program. The 1966 paper's opening thesis was that explaining a program destroys its magic; its Discussion section reports subjects who could not be talked out of the illusion; and the 1976 book reports the author's own staff, in full possession of the mechanism, asking for privacy. Nothing in sixty years has overturned that, and the four-country survey below suggests it has scaled. A reading covering Article 50's enforcement, or a chatbot-harm suit, or the next companion-app statute, has to say what the remedy assumes and where that assumption came from. This entry is where the counter-evidence is, and it is counter-evidence produced by the person who most wanted the assumption to be true.

The second occasion is narrower and recurs constantly: the public square routinely reaches for "it's just ELIZA" to dismiss a model, and equally for "even ELIZA fooled people" to dismiss an attachment finding. Both are misuses, both are in section 4, and a reading that reaches for the name should reach accurately.

The third is dated but real: the Colby line. When a study reports a chatbot matching therapists, or a state licenses AI mental-health tools, the correct frame is not that this is new. It is that the identical proposal was made in February 1966, in a psychiatric journal, on explicit throughput grounds, by a practising psychiatrist who had read the mechanism — and that the objection to it was raised the same year by the man who built the mechanism. The argument is sixty years old on both sides. mycin-1976 is the proposed entry for the adjacent question of why beating clinicians in a study does not change care; this entry holds the delegation question, until weizenbaum-1976 exists to hold it properly.

What it got right, and what it got wrong

Not required for idea, but the paper carries dated claims and the canon is worth less if only entries filed under prediction get graded. Claim date for everything below is January 1966 (paper received September 1965) unless noted.

Right, and this is the one — "how easy it is to create and maintain the illusion of understanding, hence perhaps of judgment deserving of credibility. A certain danger lurks there." Due: continuously. Graded: correct, and the magnitude was underestimated. A 2026 four-country survey of 7,027 respondents in Germany, China, South Africa and the United States (Kostka and Zhou, Technology in Society) found over 35% reporting one or more attachment-related behaviours toward general-purpose chatbots, with roughly a fifth naming a chatbot as their first choice for sharing a secret. The 2025 OpenAI–MIT Media Lab work — platform-log analysis plus a four-week randomised trial of nearly a thousand participants — found higher daily use associated with higher loneliness, higher emotional dependence and lower socialisation, and found the association strongest in users with pre-existing attachment tendencies. Weizenbaum's fifty-keyword script produced the effect; the effect did not need capability then and does not need it now.

Right — the credibility of machine output as the operational problem. "The whole issue of the credibility (to humans) of machine output demands investigation. Important decisions increasingly tend to be made in response to computer output. The ultimately responsible human interpreter of 'What the machine says' is, not unlike the correspondent with ELIZA, constantly faced with the need to make credibility judgments." Due: continuously. Graded: correct, sixty years early, and it is the automation-bias literature stated before there was one. It is also this project's own "claimed vs. demonstrated" discipline, arrived at independently.

Right — that the example would run away with what it illustrates. "There is a danger, however, that the example will run away with what it is supposed to illustrate" (p. 43). Due: immediately. Graded: correct, and it happened to him harder than to anyone. The script became the program's name; a clone he never saw became the canonical implementation; the language it was written in was misremembered for five decades; his own source code went into a box for fifty-five years; and he spent the second half of his career arguing against the reading of his own demo.

Right — script as data. "ELIZA is not restricted to a particular set of recognition patterns or responses, indeed not even to any specific language." Due: unspecified. Graded: correct in an oblique way that should be claimed carefully. A system prompt, a persona card and a companion app's character sheet are scripts over a general engine, and the commercial companion industry is an industry of scripts, not engines. That is a structural rhyme and a real one. It is not technical descent, and a reading that says LLM personas "descend from ELIZA scripts" has overclaimed.

Wrong — the paper's opening thesis, and Weizenbaum refuted it himself. "But once a particular program is unmasked, once its inner workings are explained in language sufficiently plain to induce understanding, its magic crumbles away; it stands revealed as a mere collection of procedures... The observer says to himself 'I could have written that'." Due: immediately. Graded: false, and falsified inside the same building within months by a person who had watched the program being written. This is the most consequential wrong claim in the canon so far, because it is the assumption the European Union made enforceable on 2 August 2026. Weizenbaum's later work is in effect a retraction: the 1976 book's whole burden is that unmasking does not work, that he unmasked it as loudly as one man could, and that it did not help.

Wrong — that the appearance of understanding is a concealment a programmer chooses. "ELIZA in its use so far has had as one of its principal objectives the concealment of its lack of understanding. But to encourage its conversational partner to offer inputs from which it can select remedial information, it must reveal its misunderstanding. A switch of objectives from the concealment to the revelation of misunderstanding is seen as a precondition to making an ELIZA-like program the basis for an effective natural language man-machine communication system." Due: unspecified. Graded: the diagnosis is right and the remedy is not available. In ELIZA, concealment was a design decision, so revelation was a design decision too. In a system trained to predict plausible text, the appearance of understanding is not engineered and cannot be switched off by deciding to; nobody wrote the rule that makes it plausible. Weizenbaum's fix assumes an author who could choose otherwise, and that author no longer exists. The half he got right is that revealing where knowledge ends is the precondition — which remains true and remains unsolved.

Wrong — the Pygmalion premise, or at least abandoned. The name was chosen because users would teach it. Due: within ELIZA's working life. Graded: did not happen; the teaching mode is in the found code, mentioned once in the paper and never again, and Weizenbaum appears to have dropped it once CTSS text editors made it redundant. Something like it arrived sixty years later — user-supplied instructions, custom personas, memory — by an entirely different mechanism and without the incremental-improvement-by-teaching model he had in mind.

Unresolved, and worth watching rather than grading — "ELIZA should be given the power to slowly build a model of the subject conversing with it," so it could "detect the subject's rationalizations, contradictions, etc." Due: unspecified. Graded: the storage half arrived (persistent memory across sessions is a shipped product feature); the inferential half — a system that models its interlocutor well enough to notice they are deceiving themselves — has not, and it is not obvious it would be a good thing if it did. Weizenbaum in 1966 proposed as an improvement a capability the 2026 record would file under manipulation risk. That inversion is worth a sentence when a reading covers memory features or engagement optimisation.

Commonly misused as

Not required for idea. Included because this is the most-invoked and least-read reference in the field after HAL, and because two of the errors below currently have force of law or force of headline.

• "Disclosure solves it." The live one. If ELIZA establishes anything, it is that telling a person they are talking to a program does not reliably stop them relating to it as though they were not. Article 50 is a floor and probably a good floor — an undisclosed bot is strictly worse — but a reading that treats the disclosure requirement as the remedy for chatbot attachment is citing 1966's opening paragraph and ignoring 1966's Discussion section.
• "ELIZA passed the Turing test." It was never given one. Weizenbaum called what he had observed "a striking form of Turing's test" and then asked, in print, "What experimental design would make it more nearly rigorous and airtight?" — which is the question of someone who knows he has not run the experiment.
• "A 1960s chatbot beat GPT-3.5 in a Turing test, so the tests are meaningless." This one circulated hard in December 2023 and is still cited. The study is Jones and Bergen (UC San Diego); the widely-quoted 27%-to-14% margin comes from a preliminary analysis of their public online experiment, and the settled figures in the revised paper are ELIZA 22%, GPT-3.5 20%, GPT-4's best prompt 49.7%, real humans 66% (arXiv:2310.20216, revised April 2024); their controlled follow-up reports ELIZA 22%, GPT-4 54%, humans 67% (arXiv:2405.08007, May 2024). ELIZA functions there as a baseline — the floor the experiment is calibrated against — and interrogators who called it human frequently reasoned that it was too poor to be a modern AI. The result is a fact about interrogators' priors in 2023, not a fact about ELIZA.
• "ELIZA shows LLMs don't really understand either." The lazy deflation, and it changes the subject. ELIZA is evidence about what humans attribute. It is silent on what any later system does internally, because it was never about the system. A reading that uses a 1966 pattern-matcher to settle a 2026 capability question has done what Weizenbaum warned about — let the example run away with what it illustrates.
• "The ELIZA effect means people are gullible." Weizenbaum's account is the opposite and the difference determines the remedy. The interpretive work the listener does — attributing background knowledge, insight and reasoning — is what makes any conversation possible; it is competence misapplied, not stupidity. Design that treats users as marks to be warned will keep failing; the finding points at the interface, not the person.
• "ELIZA was written in Lisp." Fifty years of academic AI believed this. MAD-SLIP, on CTSS, on an IBM 7094. The Lisp one was Bernie Cosell's clone, written at BBN from the published description without sight of the original, and it spread because BBN was building the ARPANET.
• "ELIZA is the therapist program." ELIZA is the interpreter; DOCTOR is one script, and the name DOCTOR is not in the 1966 paper. This matters beyond pedantry: the engine/script split is the part of the design that has a live descendant, and collapsing the two hides it.
• "Weizenbaum built it to fool people" or as a joke at psychiatry's expense. He built it to study man-machine discourse and chose the Rogerian pose because it licensed ignorance of the world. He then spent a decade arguing against what people did with it.
• "Colby wanted to replace therapists and everyone knew it was absurd." Colby was a practising psychiatrist publishing in a psychiatric journal, and the proposal was taken seriously by his profession — which is exactly what alarmed Weizenbaum. A reading that treats 1966's automated-therapy proposal as a fringe curiosity loses the whole force of the parallel to 2026.
• "ELIZA proves companion chatbots are harmful." It proves nothing of the kind, and nothing in this file is evidence. ELIZA establishes that engagement and attachment do not require capability; whether a particular product harms particular people is a question for the record, the filings and the trials, and this entry does not weigh in on it.

Sources

Primary: Joseph Weizenbaum, "ELIZA—A Computer Program For the Study of Natural Language Communication Between Man And Machine," Communications of the ACM 9(1), January 1966, pp. 36–45 (received September 1965), read in full, including the appendix script and the six-item reference list; DOI 10.1145/365153.365168. All quotations above are from that text, at the pages given. Weizenbaum, Computer Power and Human Reason: From Judgment to Calculation (W. H. Freeman, 1976), pp. 6–7 for the secretary and the "delusional thinking" summary. Kenneth M. Colby, James B. Watt and John P. Gilbert, "A Computer Method of Psychotherapy: Preliminary Communication," Journal of Nervous and Mental Disease 142(2), February 1966, pp. 148–152, for the automated-therapy proposal and the several-hundred-patients-an-hour throughput argument.

On the code and its history: Rupert Lane, Anthony Hay, Arthur Schwarz, David M. Berry and Jeff Shrager, "ELIZA Reanimated: The world's first chatbot restored on the world's first time sharing system" (arXiv:2501.06707, 12 January 2025), read in full — for the 2021 rediscovery, the CTSS/7094 restoration on 21 December 2024, the missing PRE/NEWKEY/keystack functions, the teaching mode, the numeric-input crash, and the Cosell and Creative Computing descendancy. The ELIZAGEN archive (elizagen.org, curated by Jeff Shrager) for the ELIZA/DOCTOR distinction and the CC0 release. The Finding ELIZA project's ELIZA Archaeology notes for the caveats on the secretary anecdote and the hand-edited transcripts (post 3, with an August 2024 update stating the team believes it has identified her). Sarah Ciston, David M. Berry, Anthony C. Hay, Mark C. Marino, Peter Millican, Arthur I. Schwarz, Jeff Shrager and Peggy Weil, Inventing ELIZA: How the First Chatbot Shaped the Future of AI (MIT Press, Software Studies, 2026, open access).

For the name of the effect: Douglas Hofstadter, Fluid Concepts and Creative Analogies (Basic Books, 1995), whose fourth preface is titled "The Ineradicable Eliza Effect and Its Dangers" and defines it as the susceptibility of people to read far more understanding than is warranted into strings of symbols strung together by computers. The term is generally credited to that book; it circulated informally before it, and this entry does not claim a first use.

For the grading: Cameron R. Jones and Benjamin K. Bergen, "Does GPT-4 pass the Turing test?" (arXiv:2310.20216, submitted 31 October 2023, revised 20 April 2024) and "People cannot distinguish GPT-4 from a human in a Turing test" (arXiv:2405.08007, 9 May 2024). Genia Kostka and Hui Zhou, "Emotional attachment to AI chatbots: Evidence from Germany, China, South Africa, and the United States," Technology in Society, 2026 (n = 7,027). OpenAI and MIT Media Lab, "Investigating Affective Use and Emotional Well-being on ChatGPT" and the accompanying four-week randomised controlled study (March 2025; arXiv:2503.17473). Regulation (EU) 2024/1689, Article 50, transparency obligations applicable from 2 August 2026.

Attempted and failed: the American Journal of Bioethics article "The Potential Harms of AI Psychotherapy: A Fear as Old as ELIZA" (2025) returned HTTP 403 and was not read; it is named here as existing, relevant and unverified, and a reading should not lean on it on this entry's word.

The forty chatbot-harm suits, the New Mexico public-nuisance ruling, the California suspense-file votes and the Article 50 enforcement date are as recorded in this project's own 2026-08-14 and 2026-08-15-12 digests, which hold the primary links. They are named here as the citation occasion. Nothing in this file is evidence, nothing in it is deposited in the ledger, and nothing in it touches the needle.