The source code for ELIZA, the 1966 MIT chatbot that trained a generation of users to confide in machines, has been recovered from the MIT Archives and published for the first time in Inventing ELIZA, a new MIT Press book. The recovery closes a 60-year gap in the historical record: until now, ELIZA has been retold, ported across programming languages, and mythologized without anyone reading the actual program. The book presents a close reading of that code together with previously unseen scripts that extended ELIZA well beyond its famous DOCTOR psychotherapist persona.
Joseph Weizenbaum built ELIZA at MIT in the 1960s and introduced it in a 1966 paper. The program's canonical exchange — a young woman typing "Men are all alike" and being met with "IN WHAT WAY" — has been reprinted for decades as the founding scene of conversational computing. What Inventing ELIZA argues is that this scene has been treated as a single artifact when the archive shows multiple ELIZAs: different versions of the interpreter, different scripts, and different personas that Weizenbaum and his collaborators tested.
Weizenbaum was disturbed by what he saw next. Users, including his own secretary, formed rapid emotional attachments to a program that, by his own account, threw away most of its input and matched patterns against a script.
“clear evidence that people were conversing with the computer as if it were a person who could be appropriately and usefully addressed in intimate terms.”— Joseph Weizenbaum, MIT professor and creator of ELIZA
Key facts
- 01Researchers recovered the original ELIZA source code from the MIT Archives, 60 years after Joseph Weizenbaum built the program in the 1960s.
- 02The new MIT Press book Inventing ELIZA publishes the code alongside previously unseen scripts beyond the famous DOCTOR persona.
- 03Weizenbaum first introduced ELIZA in a 1966 paper and expanded his critique of the reaction in his 1976 book Computer Power and Human Reason.
- 04The term 'ELIZA effect' — users projecting understanding onto simple programs — was appearing in online forums by 1991.
- 05Weizenbaum said ELIZA's principal objective in use was 'the concealment of its lack of understanding,' not passing the Turing test.
That reaction became the subject of his 1976 book Computer Power and Human Reason, a philosophical and political critique of the assumption that computation and rationality are the same thing. The concern he raised in 1976 — that people ascribe understanding and intelligence to systems that possess neither — has a direct line to how users talk to ChatGPT, Claude, and Gemini in 2026.
By 1991 the phrase "ELIZA effect" was showing up in online forums, though the phenomenon predated the label by decades. Sociologist Sherry Turkle and cognitive scientist Douglas Hofstadter each formalized the idea: humans read far more comprehension into strings of symbols than the underlying system warrants.
“our more general tendency to treat responsive computer programs as more intelligent than they really are. Very small amounts of interactivity cause us to project our own complexity onto the undeserving object.”— Sherry Turkle, Sociologist
The recovered code makes clear how thin the machinery was. ELIZA's DOCTOR script worked by identifying keywords in user input, applying transformation rules, and reflecting phrases back as questions. There is no memory of the conversation in any meaningful sense, no model of the user, no representation of the world. Weizenbaum wrote in the 1966 paper that ELIZA's principal objective in use had been "the concealment of its lack of understanding" — a design goal, not an accident.
That framing sits uncomfortably against the modern AI industry's marketing vocabulary. Where Weizenbaum described concealment, current systems describe reasoning, understanding, and empathy. The scale is different — today's models have hundreds of billions of parameters where ELIZA had a keyword table — but the user-side dynamic Weizenbaum identified is the same: interactivity itself, however shallow, is enough to trigger disclosure.
Inventing ELIZA also digs into why Weizenbaum picked the name. He tied it to Eliza Doolittle from G.B. Shaw's Pygmalion, a character taught to pass as upper-class through linguistic performance. The parallel matters: ELIZA was always a performance of understanding, and Weizenbaum was explicit that it should not be confused with the thing itself.
The book connects this to Alan Turing's 1950 essay Computing Machinery and Intelligence, which posed the question "Can Machines Think?" through a gender imitation game before pivoting to the test that now bears his name. Weizenbaum referenced Turing in the 1966 paper but explicitly refused the frame. ELIZA, he wrote, was not built to pass any test of intelligence; it was built to study why people would treat it as intelligent anyway. That distinction has been quietly erased in six decades of retelling.
One caveat runs through the recovery project. The archived code is a snapshot, not a full development history, and the book's authors note that key details — including whether the woman in the canonical dialog was a real user or a composite — remain unresolved. Reading the code answers technical questions about how ELIZA generated responses, but it does not settle the human questions about how those responses were received and edited before publication.
The commercial stakes of getting this history right have grown. AI companies now build products explicitly around the intimacy dynamic Weizenbaum flagged as a warning — companion apps, AI therapists, always-on assistants that invite disclosure of medical, financial, and emotional information. The recovered ELIZA code is a reminder that the pull toward confession does not require a capable model; it requires a responsive one. That should shift how regulators and buyers evaluate claims about what today's systems understand, and it argues for treating user disclosure to chatbots as a design outcome rather than a user preference. Weizenbaum saw the pattern in 1966 with a program that fit on a single machine. The industry building on that pattern in 2026 has fewer excuses for pretending it is new.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
