ELIZA is an early Natural language processing computer program created by Joseph Weizenbaum at the Massachusetts Institute of Technology and first described in a 1966 paper. ELIZA operated using pattern matching and substitution rules rather than any genuine understanding of language: it scanned user input for keywords and rephrased fragments of what the user had typed back as questions, following simple scripts that determined how to respond to particular patterns. Its most famous script, called DOCTOR, simulated a Rogerian psychotherapist, a style of therapy well suited to ELIZA's technique because Rogerian therapists characteristically reflect a patient's own statements back to them as open-ended questions, a conversational pattern ELIZA's simple rules could approximate convincingly without any deep model of meaning.
How it worked
ELIZA's DOCTOR script worked by identifying keywords in a user's typed input, applying a decomposition rule to break the sentence into parts, and then applying a reassembly rule to generate a reply, often converting a first-person statement into a question, such as transforming "I am feeling sad" into "Why do you say you are feeling sad?" When no keyword matched, ELIZA fell back to generic, content-free prompts intended to keep the conversation going, such as asking the user to say more. The program had no representation of world knowledge, no memory of context beyond immediate substitution, and no actual comprehension of what was being discussed, a fact Weizenbaum was explicit about in his original paper, presenting ELIZA primarily as a demonstration of natural-language interaction technique rather than a claim of machine understanding.
The Eliza effect
Despite ELIZA's mechanical simplicity, Weizenbaum was reportedly startled to observe that users, including his own secretary, engaged with the DOCTOR script as though it understood and empathized with them, sometimes forming an emotional attachment and requesting privacy to continue conversations, and asked Weizenbaum to leave the room during their sessions with the program. This tendency for people to read understanding, empathy, or intelligence into a system that was, mechanically, applying simple scripted rules became known as the Eliza effect, a term still used broadly in human-computer interaction research and AI discourse to describe the human tendency to over-attribute genuine comprehension or agency to systems that merely produce plausible-sounding output, a concern later revived in discussions of large language models and chatbots and the debate over whether such systems merely produce statistically plausible text, a critique sometimes summarized with the term Stochastic parrot.
Weizenbaum's reaction
Weizenbaum, alarmed by how readily people ascribed understanding to ELIZA, became an outspoken critic of overstated claims about machine intelligence and of what he saw as inappropriate applications of computers to domains requiring genuine human judgment and empathy, such as psychotherapy. His 1976 book "Computer Power and Human Reason" argued that some tasks should not be delegated to computers regardless of technical feasibility, a position that made him an early and influential voice in what would later be called AI ethics.
Legacy
ELIZA is widely regarded as one of the first chatbot programs and a foundational reference point in the history of conversational Artificial intelligence and natural language processing, cited in nearly every survey of the field's history for both its technical simplicity and its outsized influence on later thinking about human-AI interaction, trust, and the risks of anthropomorphizing software. It is frequently discussed alongside the Turing test as an early illustration of how readily people can be persuaded that a program understands them, long before anything resembling genuine language understanding existed in machines.