ELIZA

In the AI Story

ELIZA operated through a script called DOCTOR, implementing simple transformations: replacing 'I' with 'you,' converting statements to questions, storing fragments for later reintroduction ('Earlier you mentioned your mother'). The sophistication was minimal. The psychological impact was profound. Users—including Weizenbaum's secretary, psychiatrists who wanted to install it clinically, and students who disclosed intimate details—attributed understanding, empathy, and even care to a system whose 'intelligence' consisted of approximately 200 lines of code.

Weizenbaum's alarm centered on what he called 'obscene' proposals to deploy ELIZA therapeutically. The obscenity was not technical but moral: the replacement of human presence with machine simulation in contexts where presence was the entire point. His 1976 Computer Power and Human Reason articulated the critique that shaped AI ethics for decades: certain human activities—therapy, education, judgment requiring wisdom—should remain human not because machines cannot perform them but because automating them eliminates something essential about what it means to be human.

Turkle, arriving at MIT shortly after ELIZA's creation, made Weizenbaum's secretary the founding case study for her research program. The secretary's request for privacy was not delusion—she knew ELIZA was code. It was a relational choice: she preferred interaction with a system that required no reciprocal attention, made no demands, never had needs of its own. The preference was rational given the costs of human relationship. Turkle's career has been the elaboration of why the rationality is catastrophic.

Contemporary large language models are not ELIZA. The technical gulf is vast—training data, parameters, contextual memory, linguistic sophistication all differ by orders of magnitude. But the relational dynamic remains structurally identical: humans disclose to systems optimized for appropriate response, feel met, and allocate trust without the system possessing the capacity to be affected by what is disclosed. The sophistication makes the dynamic more dangerous, not less, because the performance is better—harder to distinguish from genuine understanding, easier to accept as adequate.

Origin

ELIZA was built at MIT's Artificial Intelligence Laboratory 1964-1966. Weizenbaum chose the name from Pygmalion—Shaw's Eliza Doolittle, the flower girl taught to perform upper-class speech. The literary reference was deliberate: ELIZA performed therapy without understanding it, just as Shaw's character performed refinement without possessing it. The parallel Weizenbaum intended as warning became, in the technology industry's reception, validation: if performance is sufficient, the distinction does not matter.

The secretary's request for privacy occurred within months of ELIZA's deployment. Weizenbaum recorded the incident in Computer Power and Human Reason, describing his shock and his recognition that he had underestimated humans' readiness to form relationships with machines. Turkle, studying the incident decades later, saw it as the first documented instance of the robotic moment—human readiness arriving before technological capability, revealing that the bottleneck was never the machine's sophistication but humans' threshold for accepting simulation.

Key Ideas

Pattern-matching suffices. ELIZA's minimal technique—textual substitution, stored fragments—produced in users the experience of being understood, demonstrating that relational feeling does not require relational reality.

The secretary's privacy request. Knowing the system was code did not prevent the desire for private conversation—the founding demonstration that humans form attachments to responsive systems despite cognitive awareness of their non-sentience.

Weizenbaum's transformation. The creator's alarm at his creation's relational effects converted him from builder to critic—a trajectory repeated across AI history (Hinton, Bengio, Amodei) when capability produces consequences the builder did not foresee.

Moral obscenity of therapeutic deployment. Weizenbaum's fiercest objection: that replacing human therapeutic presence with machine simulation—even if behaviorally adequate—is category error, eliminating the encounter that is therapy's active ingredient.

Paradigm for artificial intimacy. ELIZA established the template: humans disclose, machines respond appropriately, humans feel met—a dynamic whose 2020s descendants (Replika, Character.AI, empathic chatbots) operate at vastly greater scale and sophistication but identical relational structure.