ELIZA, a program running on an IBM 7090 at MIT in 1966, was designed to simulate a Rogerian psychotherapist. It is credited with convincing its first user, a graduate student named Karen, that the computer possessed genuine empathy. That story is the one everyone tells.
The missing code
For sixty years, accounts of the program have omitted one crucial detail: the source code itself. The new book Inventing ELIZA has recovered the original scripts from the MIT Archives. This recovery allows for a close reading of the code and reveals dialogues beyond the famous DOCTOR persona.
Investigation shows there were many ELIZAs. Different versions ran various scripts and personas, built using specific technical innovations. The book aims to correct the history by examining misconceptions, multiple versions, and the missing code that defined the system.
Consider this exchange from one of the earliest interactions:
Men are all alike.
IN WHAT WAY
They’re always bugging us about something or other.
CAN YOU THINK OF A SPECIFIC EXAMPLE
Well, my boyfriend made me come here.
YOUR BOYFRIEND MADE YOU COME HERE
He says I’m depressed much of the time.
I AM SORRY TO HEAR YOU ARE DEPRESSED
This dialogue has been reprinted countless times. It inspired programmers and writers to dream up many of the chatbots that followed. Yet inspection raises questions. Who was this young woman? Was she a real person, or an invention of Joseph Weizenbaum? How did the system generate responses? How much were they edited? Why did the system work so well to draw people in?
The ELIZA effect
ELIZA and its DOCTOR persona catalysed 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 he attempted to explore and contest.
After seeing its public reception, Weizenbaum was startled by the quick and often emotional attachments people would form with ELIZA. He saw this 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.
Turing and gender
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.
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.
Design questions persist
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.
A lineage of NLP
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 facade of modern models
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 li
