How ELIZA, the 1960s Chatbot, Predicted Today's AI Attachment Problem

ELIZA has dominated historical accounts as the earliest example of what is now called a chatbot, a program that could converse as an automated psychologist. The famous dialog—in which a user types "Men are all alike," and ELIZA responds "IN WHAT WAY"—has been reprinted countless times and inspired generations of programmers. However, the source code for ELIZA itself has remained largely inaccessible. A new book, Inventing ELIZA, recovers this source code from the MIT Archives and offers the first close reading of the program's code, along with newly discovered dialogs for ELIZA scripts beyond the popular "DOCTOR" persona. The investigation reveals multiple versions of ELIZA, each designed to run different scripts and personas using a series of technical innovations.

Weizenbaum created ELIZA not to demonstrate machine intelligence but to explore the psychological factors that lead humans to attribute intelligence to machines. In his 1966 paper introducing ELIZA, Weizenbaum explicitly distanced his creation from claims of understanding, writing that "the crucial test of understanding is not the subject's ability to continue a conversation, but to draw valid conclusions." He noted that "ELIZA throws away [most] of its inputs" and that the program's principal objective was "the concealment of its lack of understanding." Yet when ELIZA was released, people's reactions surprised even its creator. 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." This tendency became known as the "ELIZA effect," a term defined by sociologist Sherry Turkle 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."

The naming of the system was itself a statement about performance and identity. Weizenbaum chose the name "Eliza" after Eliza Doolittle from G.B. Shaw's Pygmalion, a working-class woman taught to pass as upper-class through linguistic transformation. As Weizenbaum explained, "like Miss Doolittle, it was never quite clear whether or not it became smarter." This reference connected ELIZA to Turing's earlier imitation game, in which gender performance became a metaphor for machine imitation. The ELIZA system performed personas—most famously the "DOCTOR"—through scripted and repetitive linguistic patterns, without possessing humanlike understanding, much as Eliza Doolittle performed class through speech acts. What remains striking in historical accounts is that the women who appeared in ELIZA's published dialogs were never named, while they conversed with a therapist called DOCTOR—a title that, in the 1960s, carried masculine associations. This erasure embedded gendered assumptions into the software itself.

ELIZA's influence on subsequent technologies has been vast. It helped launch the field of computational agents and intersected with innovations in string processing, text synthesis, entity recognition, and sentiment analysis. It emerged alongside machine translation, semantic networks, and speech recognition—techniques that eventually coalesced into natural language processing (NLP), the computational field that deals with how computers parse, interact with, and output human language. Today, contemporary large language models like OpenAI's ChatGPT retain a resemblance to Weizenbaum's original chatbot interface. Yet the machinery behind these systems remains opaque to most users. As the authors note, the "enticing facades obfuscate the machinery that often includes a combination of statistical predictions, rule-based procedures, and human labor disguised as machine labor," leaving users with "limited opportunity to distinguish hype from substance." Weizenbaum warned of the dangers this obfuscation creates, arguing that removing language from its social contexts and treating it as abstract computational symbols "can be dehumanizing," risking "rights violations, privacy breaches, exploitation, displacement, and discrimination." The recovery of ELIZA's code and history shows that these concerns are not new—they are foundational to how AI systems have been designed and deployed for more than sixty years.

ELIZA emerged in the 1960s at a moment when Alan Turing's foundational 1950 essay "Computing Machinery and Intelligence" had posed the question "Can Machines Think?" through his famous imitation game. Weizenbaum's choice to name his program after Eliza Doolittle—the character from Shaw's Pygmalion who learns to perform upper-class speech despite not changing her nature—was deliberate: it signaled that the system was performing an identity through language, not achieving genuine understanding. This design choice, grounded in questions of performativity and identity, proved prophetic. When users encountered ELIZA's DOCTOR persona, they attributed empathy and understanding to a system that merely followed scripts. Weizenbaum was astonished by the emotional attachments people formed, observing that they conversed with the machine "as if it were a person who could be appropriately and usefully addressed in intimate terms."

The recovery of ELIZA's original code and dialogs reveals a pattern that has been largely obscured in historical accounts: the women who appeared in published conversations with ELIZA remained unnamed, while the system itself performed a gendered identity as an unnamed male "DOCTOR." This erasure and gendering of the interaction points to deeper assumptions about embodiment, labor, and what gets hidden inside a computational system. Weizenbaum warned in his 1976 book Computer Power and Human Reason that removing language from its social contexts and treating it as abstract computational symbols could be dehumanizing—a concern that resonates acutely today. Contemporary large language models like ChatGPT inherit ELIZA's chatbot interface and its capacity to elicit emotional investment, yet users often have even less visibility into how the system works, where the training data comes from, or what human labor is embedded in its outputs.

No comments yet. Be the first to share your thoughts!