ELIZA

Joseph Weizenbaum developed ELIZA at the Massachusetts Institute of Technology (MIT) between 1964 and 1967, in the programming language MAD-SLIP, on an IBM 7094 computer running the CTSS operating system. Weizenbaum named the program after Eliza Doolittle, the character in George Bernard Shaw's play "Pygmalion" (on which "My Fair Lady" is based). Just as the fictional character was taught to speak with an upper-class accent, ELIZA the program was designed to improve gradually and seem more convincing as users contributed more conversation patterns to it.

ELIZA did not "understand" anything in the real sense. It identified keywords in the user's sentence, ranked them by priority, and then applied "decomposition and reassembly" rules to produce a response, usually a rephrasing of the user's words as a question. The best-known script, "DOCTOR," imitated a psychotherapist of the Rogerian school, a therapeutic approach based on non-judgmental mirroring of the patient's words back to them. Choosing it let Weizenbaum "sidestep the problem" of equipping the program with real knowledge about the world.

Despite its technical simplicity, many users described a deep emotional connection. Foremost among them was Weizenbaum's own secretary, who asked him to leave the room so she could "talk" with ELIZA in private. Weizenbaum was stunned, and later wrote: "I had not realized ... that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people." The phenomenon later came to be called the "ELIZA effect." It is early evidence of how easily people tend to "humanize" systems that only imitate conversation, a subject that is all the more relevant in today's era of sophisticated chatbots and large language models.

That shock turned Weizenbaum from ELIZA's creator into a sharp critic of AI. In 1976 he published the book "Computer Power and Human Reason," which warned against an overly anthropomorphic attitude toward computers. ELIZA's original source code was considered lost for more than 50 years, until it was found in the MIT archives and republished in 2021. This lets researchers today examine exactly how the system that gave rise to the question "does a machine understand me?" worked on the inside.