In the 60 years that ELIZA has been influencing computation and culture, conventional accounts portray it as the earliest example of what we now call chatbots, one that could converse as an automated psychologist. The deceptively simple program is known for “fooling” even the secretary who watched MIT professor Joseph Weizenbaum create it. That’s how the story goes.
However, in all those accounts—even after all its adaptations across programming languages and research fields, in classrooms and popular culture—one essential piece of the story has been missing: the source code for the ELIZA program itself. Our new book, Inventing ELIZA, recovers this source code from the MIT Archives, offering for the first time a close reading and discussion of that code along with newly uncovered dialogs for ELIZA scripts beyond its popular “DOCTOR” persona.
This investigation revealed many ELIZAs: in its different program versions, designed to run a wide variety of scripts or personas, built using a series of technical innovations. Inventing ELIZA seeks to correct and to complicate ELIZA’s history and influence by exploring the misconceptions, multiple versions, and missing code of ELIZA. In this excerpt from the book, we examine one of ELIZA's earliest interactions, how it laid the groundwork for human relationships with computers for decades to come, and how the complex program continues to speak to the unrestrained drive of today's AI industry.
They’re always bugging us about something or other. Well, my boyfriend made me come here.
He says I’m depressed much of the time.
That dialog has been reprinted countless times and has inspired programmers and writers to dream up many of the chatbots that followed. Yet the closer one inspects that dialog, the more questions arise: Who was this young woman? Was she a real person, or is she the invention of ELIZA creator Joseph Weizenbaum? How exactly did the ELIZA system generate its responses, and how much were they edited? Why did the system work so well to draw people in?
ELIZA, and her “DOCTOR” persona, helped catalyze a mode of thought and an anxiety about people’s relationships with computers. Weizenbaum explored this in his 1976 book Computer Power and Human Reason, invoking philosophical, social, and political critiques. The unique machine interaction presented by his program revealed how new forms of human-computer relation would have profound effects that he attempted to explore and to contest. After seeing its public reception, Weizenbaum was startled by the quick and often emotional attachments people would form with ELIZA, which he saw as “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.” The tendency to attribute empathy and invest private feelings into a computer puzzled Weizenbaum. He was concerned by the extent to which people associated rationality with computation, and ascribed understanding and intelligence to computer systems where none existed.
This tendency became known as the “ELIZA effect.” By 1991 the term was appearing in online forums, but its use predated that appearance by decades. Sociologist Sherry Turkle defines “the ELIZA effect” as “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.” Cognitive and computer scientist Douglas Hofstadter describes it as “the susceptibility of people to read far more understanding than is warranted into strings of symbols—especially words—strung together by computers,” which applies easily to generative AI systems today.
To understand the power and provocation of ELIZA, we can look to the infamous challenge formulated by computer scientist Alan Turing in the essay “Computing Machinery and Intelligence,” in which Turing posed the question “Can Machines Think?” Turing premised his thought experiment on a parlor game—not about technology but about gender: A man and a woman are hidden in a separate room and an interrogator tries to identify who is which gender by asking a series of questions. The man tries to mislead the interrogator, pretending to be a woman, while the woman tries to convince the interrogator of the “correct” answer. That is, both of them claim they are the “real” woman, a challenge to essentialist notions of gender.
In his revision of this game, Turing replaces the original gender question with what is now called the Turing test. In this challenge a machine pretends to be a man, rather than a man pretending to be a woman. More than a mere opening gambit, Turing’s choice of the initial gender imitation game has ensured that artificial intelligence would remain intertwined with questions of gender and identity. In this sense imitation, drag, and gender deconstruction laid the groundwork for AI and the performance of intellect. Weizenbaum’s ELIZA picks up where Turing left off, not least with the opening line of the dialog: “Men are all alike.”
Nonetheless, even though Weizenbaum references Turing’s imitation game in his 1966 paper introducing ELIZA, he explicitly distances his creation from any claims of intelligence: This assessment shows that ELIZA was never intended to pass the Turing test, but rather to explore the psychological factors that might lead humans to misinterpret its capabilities.
In that tradition, ELIZA continues to play with the notion of performative identity, as a vehicle for personas made out of responses that stick to the script, literally and figuratively. By naming the system after Eliza Doolittle—a working-class female character who is taught to pass as an upper-class white woman—Weizenbaum continues Turing’s provocation about the performance of identity. “I chose the name ‘Eliza,’” Weizenbaum says, “because, like G.B. Shaw’s Eliza Doolittle of Pygmalion fame, the program could be taught to ‘speak’ increasingly well, although, also like Miss Doolittle, it was never quite clear whether or not it became smarter.”
As Shaw’s Eliza performs race and ethnicity, class, sexuality, and gender through linguistic transformation, the ELIZA system performs an assumed persona through scripted and repetitive linguistic patterns and transformations, but without possessing humanlike understanding. Feminist philosopher Judith Butler’s theories of gender performativity offer a possible framework to think about this. Butler argues that gender and sexuality are not innate but are performed through repeated acts. They are iterated. Just as Eliza Doolittle challenges assumptions of class by performing speech that help her pass as among the upper classes, Weizenbaum’s system performs personas like DOCTOR through speech acts, or code acts, and sample dialogs that perform gendered, classed, and racialized identities.
Joseph Weizenbaum, creator of ELIZAPhotograph: Wolfgang Kunz/Getty Images
In this way, ELIZA both reinforces and challenges assumptions about gendered and embodied communication. It is interesting to note that in the published dialogs and popular stories about ELIZA, the women who appear alongside the program remain unnamed. They are conversing with a therapist called DOCTOR, a name that, though it does not carry gender markers today, would have sounded like a man’s title in the 1960s. Therefore, stories that suggest that these women confessed their secrets to this artificial doctor carry a gendered message, including the fantasy that one can be disembodied. In this way, the ELIZA system becomes a site where questions of identity, performativity, and embodiment play out across historical algorithms in ways that have implications for contemporary AI systems.
As norms, values, and ideology become embedded within algorithmic forms, exploring details of a software object like ELIZA reveals culturally and historically specific assumptions about what software can and should be, including ideas about what technology is for and how it has developed. While ELIZA might look unsophisticated to a contemporary audience, it was already addressing, in the 1960s, many of the design questions that persist in the systems we use today. Fundamentally, these systems ask how humans and machines should interact, by what means communication can be represented computationally, and to what extent machines can and should be imbued with an ability to influence users.
ELIZA is a common touchpoint not only because it was one of the first chatbots and helped launch a field of computational agents, but also because it intersects with so many of the computing innovations that followed. Alongside developments for string processing and text analysis, it influenced research on text synthesis, entity recognition, and sentiment analysis. It emerged alongside research in machine translation, semantic networks, speech recognition, speech synthesis, and other techniques that developed into the group of tasks we now call natural language processing (NLP), the area of computation that deals with how computers can parse, interact with, process, and output languages that are used by people as opposed to programming languages used by computers. In practice, these tasks are often combined to create automated agents and many other kinds of systems.
The reappearance of ELIZA-like chatbot interfaces in contemporary large language models shows that early software genealogies can help us understand new technologies. ELIZA is a useful foil for emerging models because though much has changed, much more remains the same. The history of NLP emerges from and overlaps the era of ELIZA and early AI work at institutions like MIT. In that field, trends have come and gone: Although different styles—syntactic, semantic, statistical, stochastic—were popular at different times, all continued to develop in parallel. Even now, as the latest large language models have amazed observers with the seeming intelligence of text output, the actual power of contemporary systems like OpenAI’s ChatGPT (generative pretrained transformer) are disguised behind their chatbot interfaces, which retain a resemblance to Weizenbaum’s original. These enticing facades obfuscate the machinery that often includes a combination of statistical predictions, rule-based procedures, and human labor disguised as machine labor. From the perspective of the user, this leaves limited opportunity to distinguish hype from substance, to understand how each system operates, and to see why it produces certain outputs.
Weizenbaum warned of the many issues this obfuscation could create. It can, for example, lead to exploitation as humans are replaced, harmed, or treated unfairly by these systems. As Weizenbaum says, “an individual is dehumanized whenever he is treated as less than a whole person. The various forms of human and social engineering … do just that, in that they circumvent all human contexts, especially those that give real meaning to human language.” Weizenbaum argued that removing language from its social contexts and treating it as a set of abstract concepts in a computational system can be dehumanizing. It risks ignoring the multiple meanings inherent to language, which are impossible to capture fully in AI systems and which can result in direct harm, rights violations, privacy breaches, exploitation, displacement, and discrimination. That is why it is essential to consider broader ethical and social impacts when designing, deploying, and using automated systems.
Emerging large language models support a new class of knowledge-powered systems. Under the surface, the chatbot interface runs on social labor. Its machinery is supported by literally millions of traces of humans’ writings and conversations, siphoned into datasets usually without creators’ awareness or consent. That labor enables automated cultural production that is managed, controlled, monitored, disaggregated, and reaggregated on demand. This abstracts human cultural production through software, treating it as a standing reserve for a computational system (in a similar fashion to electricity or a water supply, yet privatized and monopolized).
As Langdon Winner observed in 1977, it is possible “for inanimate instruments to perform their own work ‘at the word of a command or by intelligent anticipation,’ that is, by a computer program. This development has led to conjecture that the perfection of industrial technology will eventually liberate mankind from toil,” a common story that OpenAI, Anthropic, and other AI companies repeat today. Yet computation is always serviced by human labor—even as chatbots currently manage customer queries and problems, help with homework and replace teaching assistants, converse as companions and counselors, and entertain and entangle with notions of identity and human exceptionalism. They also help produce a deluge of AI slop that is cannibalizing its source materials and the planet’s natural resources. The idea that ELIZA might inspire computational systems that create cognitive factories and AI slop would likely have horrified Weizenbaum. It is, in a sense, the realization of cybernetic systems that attempt to closely couple machines and humans into tightly linked feedback loops—exactly the threat that Weizenbaum warned about in Computer Power and Human Reason.
In that book and other post-ELIZA writings, Weizenbaum speaks out as an early critic of what we now call the tech sector and its mindset of exponential acceleration, regardless of the exploitative relationships it may be encoding and the social and political consequences of abstracting computational systems. His own relationship to technology shifted as he saw the impact of ELIZA and the ways automated systems were being scaled up and leveraged for power, trends that have continued with a wider public fascination and unease with the emergence of “intelligent” machines.
Excerpted from Inventing Eliza: How the First Chatbot Shaped the Future of AI. Copyright © 2026 by Sarah Ciston, David M. Berry, Anthony C. Hay, Mark C. Marino, Peter Millican, Jeff Shrager, Arthur I. Schwarz, and Peggy Weil. Used with permission of the publisher, The MIT Press. All rights reserved.