Skip to content

Joseph Weizenbaum and ELIZA

Abstract

In 1966 an MIT professor published a pattern-matching program that imitated a psychotherapist, and then spent the remaining forty years of his life alarmed by what happened next. Joseph Weizenbaum (1923–2008), a Jewish refugee from Nazi Berlin who helped computerize American banking before joining MIT, built ELIZA as a demonstration that machine “conversation” could be shallow trickery. Instead, users (including his own secretary) confided in it. His 1976 book Computer Power and Human Reason turned that shock into one of the founding texts of computer ethics: computers can decide, he argued, but only humans can choose. The field mostly shrugged; Germany, where he returned in his final years, named a research institute after him.

From Berlin to Detroit

Joseph Weizenbaum was born on January 8, 1923 in Berlin to Jewish parents. In January 1936 the family escaped Nazi Germany and settled in Detroit, a rupture Weizenbaum later said left him permanently an outsider, and permanently suspicious of societies that mistake technical order for moral order. He began studying mathematics at Wayne State University in 1941, interrupted by wartime service as a U.S. Army Air Corps meteorologist (1942–1946), and finished a B.S. (1948) and M.S. (1950) in mathematics. At Wayne State around 1952 he worked as a research assistant on analog computers and assisted Harry Huskey on the design that became the Bendix G-15, one of the first small digital computers.

In 1955 or 1956 (sources differ) he joined General Electric to work on ERMA (Electronic Recording Machine, Accounting), the Bank of America project that automated check processing with magnetic-ink character recognition and pushed banking into the computer age. Weizenbaum, the later critic of computerized society, was thus one of its early engineers; he knew exactly what he was criticizing.

MIT and SLIP

His list-processing system SLIP (“Symmetric List Processor”) brought him to MIT in 1963, in the middle of Project MAC’s time-sharing revolution (see Fernando Corbató and Time-Sharing). He was a full professor by 1970. A February 1962 Datamation piece he wrote carried a title that reads today like a thesis statement for his whole career: “How to Make a Computer Appear Intelligent”; it described a program for gomoku, the Japanese five-in-a-row game, played here on a 19-by-19 board. It did not search ahead in the manner of Shannon’s minimax; it only scored the position on the board and picked the move that most cut the opponent’s chances. A GE colleague coded it, it beat beginners even when they moved first, it lost to players who learned its habits, and Weizenbaum himself never beat it. The Heinz Nixdorf MuseumsForum counts it as the first gomoku program, and the title already showed his doubts about what such success proved.

ELIZA: The Trick That Worked Too Well

Between 1964 and 1966 Weizenbaum wrote ELIZA, published as “ELIZA — A Computer Program For the Study of Natural Language Communication Between Man And Machine” in Communications of the ACM, January 1966. Named for Eliza Doolittle in Pygmalion (a character taught to appear refined) the program ran on MIT’s IBM 7094 under CTSS and was written in MAD-SLIP, not Lisp as often assumed. ELIZA is often called the first program to produce emotional, natural-language text; it was not. Christopher Strachey’s Love Letters, written on the Ferranti Mark I in 1952, had done the same thing by the same template-and-word-list method fourteen years earlier.

The mechanism was deliberately shallow: scan the input for keywords ranked by precedence, apply decomposition rules to break the sentence apart, then reassembly rules to transform it into a response (“I am unhappy” → “How long have you been unhappy?”). Scripts defined the keywords; the famous DOCTOR script imitated a Rogerian psychotherapist, a persona chosen precisely because reflecting the patient’s statements back is plausible conversational behavior that requires no knowledge of the world.

A CBS television crew filmed Weizenbaum with ELIZA at MIT around the turn of 1966 to 1967. What stunned its author was the users: people knew it was a program and confided in it anyway. His secretary, who had watched him build it, asked him to leave the room so she and ELIZA could talk privately. When he proposed logging conversations for analysis, he was accused of spying. “I had not realized,” he wrote, “that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people.” The phenomenon now has a name (the ELIZA effect, the human tendency to read understanding into systems that merely manipulate symbols) and it is arguably the single most commercially important fact about conversational AI (see also The Natural Language Processing Revolution). It also set the terms for affective computing, the field that later tried to run the exchange in the other direction and have the machine read the human.

The psychiatrist Kenneth Colby, who had collaborated with Weizenbaum, went on to build PARRY (1972), a simulation of a paranoid patient (PARRY and ELIZA were even connected for a machine-to-machine “session” at the first ICCC conference in 1972) and to argue seriously that chatbot therapy could substitute for scarce human therapists. For Weizenbaum this was the moment the joke stopped being funny: the point of ELIZA had been that the empathy was fake, and here were experts proposing to administer the fake at scale.

Computer Power and Human Reason

He had already warned German readers in “Alptraum Computer” in Die Zeit in January 1972. His full answer, written during a sabbatical in 1974 and 1975, was Computer Power and Human Reason: From Judgment to Calculation (1976), published in German in 1977 as Die Macht der Computer und die Ohnmacht der Vernunft: his only monograph, and one of the founding documents of computer ethics. Its central distinction: deciding is a computational activity; choosing is not. Decision (selecting among options by rule) can be delegated to machines. Choice rests on judgment, values, and what he insisted on calling wisdom, and delegating that is not a technical error but a moral abdication. Some things a computer could do, he argued, it ought not do; he singled out automated psychotherapy and battlefield decision-making. The book also contains his acid portrait of the obsessive “compulsive programmer,” an early sighting of hacker pathology from inside the MIT AI culture he was breaking with.

The AI establishment’s response was largely hostile (John McCarthy panned the book as moralistic and anti-science) and Weizenbaum spent the rest of his career as his field’s most prominent internal dissident, arguing against Vietnam-era military funding of computing and, later, against treating human minds as information processors. A 1978 television debate set him against the British AI researcher Donald Michie. He received the CPSR Norbert Wiener Award for Social and Professional Responsibility in 1988; the line of critique he opened runs directly to modern AI ethics and algorithmic-bias research.

Return to Berlin

Weizenbaum retired from MIT in 1988 and in 1996 moved back to Berlin, an emigrant’s return that made him a fixture of German public debate about the information society, where his blend of technical authority and moral skepticism found a far warmer audience than at home (see Germany’s Computing Pioneers). He died on March 5, 2008 in Gröben near Berlin and is buried in Berlin’s Weißensee Jewish cemetery. In 2017 Germany founded the Weizenbaum Institute for the Networked Society in Berlin, a federally funded research institute for the digital society, named for the man who had warned about it.

The Code, Rediscovered

For decades ELIZA survived only through reimplementations (a BASIC version published in Creative Computing in 1977 spread it to a generation of home computers). A printout of the original MAD-SLIP source code, with an early version of the DOCTOR script, was found among Weizenbaum’s papers at MIT by Jeff Shrager and archivist Myles Crowley in 2021 and open-sourced with the estate’s permission. On 21 December 2024 a team ran it again under CTSS on an emulated IBM 7094, for the first time in its original environment in almost sixty years; the restoration turned up a little-known “teaching mode” that let users add rules by typing “+”.

⚠️ Dead End That Wasn’t: The Machine Therapist

Weizenbaum won the argument in his century and lost it in this one. Colby’s automated-therapy program looked like a dead end for decades: PARRY was a curiosity, 1980s “therapy” software a gimmick. But the ELIZA effect never stopped working: users bonded with fifty lines of BASIC in 1980, and they bond far harder with LLM companions now. Therapy and companionship chatbots are today a real industry with real clinical trials, and real documented harms, from emotional dependence to chatbots mishandling users in crisis. Weizenbaum’s warning was not that machine empathy wouldn’t work; it was that simulated empathy works on humans regardless of whether anything is behind it, which makes it precisely the kind of power he thought should not be exercised lightly. The dead end came back to life exactly as he feared it would.

Fun Fact

ELIZA is one of the few programs with an entry in both the history of AI and the history of AI criticism: its author’s disillusionment with his own creation’s reception produced one of the first books of computer ethics. The moment his secretary asked him to leave the room may be computing’s most consequential coffee-break anecdote: see Fun Fact: ELIZA Horrified Its Creator.

📚 Sources