Researchers behind Inventing ELIZA say they have recovered source code for Joseph Weizenbaum’s ELIZA from the MIT Archives, filling a gap in the history of one of computing’s most mythologized programs. The find matters because ELIZA is still treated as the family ancestor of ChatGPT-style systems, usually through a tidy story about a 1960s chatbot that sounded like a therapist and made people open up.
The book, by Sarah Ciston, David M. Berry, Anthony C. Hay, Mark C. Marino, Peter Millican, Jeff Shrager, Arthur I. Schwarz, and Peggy Weil, argues that the standard version leaves out too much. According to the authors, the archive material shows multiple ELIZA versions, not a single canonical bot, and includes scripts beyond the famous DOCTOR persona.
DOCTOR is the version most people know: a program that answers users with therapist-like prompts, often by reflecting their words back at them. The familiar sample exchange begins with a user saying, “Men are all alike,” and the program replying, “IN WHAT WAY.” The authors ask basic questions that usually get skipped: who the user was, whether the dialogue was real or composed by Weizenbaum, how much editing occurred, and how the program produced the exchange.
A machine built to expose misunderstanding
Weizenbaum did not present ELIZA as proof that a computer understood language. In his 1966 paper, cited by the book, he wrote that ELIZA’s work included the “concealment” of its lack of understanding. The program, he said, discarded most of what users typed rather than storing enough context to draw real conclusions.
That is the mechanism and the trick. ELIZA used scripted personas and text transformations to keep a conversation going while avoiding the hard work of comprehension. Users supplied the meaning. The machine supplied a plausible surface.
Weizenbaum later treated the public reaction as a warning. In his 1976 book Computer Power and Human Reason, he argued that people were willing to address computers in intimate terms and attribute understanding where none existed. That tendency later became known as the “ELIZA effect.” Sociologist Sherry Turkle defines it as people treating responsive programs as smarter than they are, while Douglas Hofstadter describes it as reading too much understanding into computer-generated strings of words.
The old interface problem did not go away
The authors connect ELIZA’s design to Alan Turing’s imitation game and to questions of gender, identity, and performance. Weizenbaum named the system after Eliza Doolittle from George Bernard Shaw’s Pygmalion, saying the program could be taught to “speak” better without it being clear whether it became smarter. The book argues that the DOCTOR script also carried gendered assumptions, especially in stories about unnamed women confiding in an artificial doctor.
The book’s sharper contemporary claim is that modern large language models inherit part of ELIZA’s interface problem. The authors say systems such as OpenAI’s ChatGPT present a chat surface that can hide the machinery underneath: statistical prediction, rules, and human labor. From a user’s seat, that surface can make it hard to tell capability from performance. Silicon Valley keeps rediscovering theater and calling it product design.
The authors also argue that today’s chatbot systems depend on vast stores of human writing and conversation, often collected without creators’ awareness or consent. They place that critique in Weizenbaum’s line of concern: when language is stripped from social context and treated as raw material for computation, automated systems can produce harm through privacy breaches, discrimination, exploitation, displacement, and other failures.
ELIZA looked primitive by modern standards, but the archive recovery suggests it was already wrestling with problems that AI companies still prefer to hand-wave: what the system knows, what the interface makes users believe, and who pays the human cost when software performs understanding.
This story draws on original reporting from WIRED.
