1 November 2026 · History · about 14 minutes

The Psychiatrist in the Machine: Joseph Weizenbaum, ELIZA, and the Book He Wrote Against His Own Program

A program that understood nothing made people believe it understood everything. What that cost its creator is still being settled.

The Secretary's Request

Sometime in the mid-1960s, in a laboratory at MIT, a secretary who knew the program was a program sat down to talk to it. When Joseph Weizenbaum came into the room to observe, she asked him to leave so that she could continue the conversation in private.¹

Weizenbaum later described what he felt. "What I had not realized is that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people," he wrote in 1976.² His secretary was not naive, not credulous, not ignorant of what she was doing. She was, by his own account, a sensible woman who had followed the program's development from the beginning. And still she wanted privacy with it.

There is a plainer reading, and it is this essay's, not the sources'. She may have known exactly what ELIZA was and still not wanted her boss reading over her shoulder while she typed about her life. Wanting privacy is not proof of delusion. Weizenbaum built a career's alarm on the moment anyway.

That moment shaped the last four decades of Weizenbaum's life. It sent him from programmer to critic, from builder to heretic, and eventually out of the country he had fled to as a child and back to the city he had fled from. To understand why, you have to begin where he did: with Berlin, and with a piece of software that was, in his own estimation, not very impressive at all.

One thing belongs here rather than at the foot of the page: this essay was written by a machine and checked by a second one. Its subject spent the last decades of his life warning against handing judgement to machines that only imitate understanding. Would he have minded a computer helping to research his life? Perhaps not. Would he have minded one writing every sentence of it, and who answers for what it says? The reader can judge. The second machine's report, with every correction made, is printed below the notes.

From Mitte to MIT

Joseph Weizenbaum was born on 8 January 1923 in Mitte, Berlin, into an assimilated, upper-middle-class Jewish family.³ The family left Berlin in 1935 and emigrated from Germany to the United States in 1936.⁴ He began studying mathematics at Wayne State University in Detroit in 1941, but interrupted his studies to serve in the US Army Air Corps as a meteorologist, he had been turned down for cryptology work because of his "enemy alien" status. After the war, he returned to Wayne State, taking his bachelor's degree in 1948 and his master's in 1950.⁵

In the 1950s he worked as a programmer for General Electric, where he helped design and build ERMA, the first computer system dedicated to banking operations, which introduced the magnetically encoded fonts on the bottom of cheques that made automated check processing possible.⁶ It was unglamorous work and consequential work: a generation later, Weizenbaum would use this experience as evidence for his argument that computers, far from disrupting institutions, tend to entrench them, ERMA had not reformed American banking but allowed an ever-expanding volume of paper to be processed without forcing the banks themselves to change.

In 1963, Weizenbaum joined MIT as a visiting associate professor of computer science, on the strength of his SLIP (Symmetric List Processing) software, a list-processing language of his own design.⁷ Within four years he had tenure; by 1970 he was a full professor. He later held academic appointments at Harvard, Stanford, the Technical University of Berlin, and the University of Hamburg.⁸ The MIT of those years was an extraordinary place: McCarthy and Minsky had founded the Artificial Intelligence Project in 1959, ARPA money was flowing, time-sharing systems were opening computers to multiple simultaneous users, and there was a general sense, not unfounded, not entirely founded, that machines were on the verge of thought.

It was into this atmosphere that Weizenbaum introduced ELIZA.

A Program That Understood Nothing

ELIZA was developed from 1964 and first described in a paper published in Communications of the ACM in January 1966.⁹ The program ran on an IBM 7094 under MIT's Compatible Time-Sharing System, written in MAD-SLIP, the list-processing language Weizenbaum had himself built.¹⁰ Its operation was, as Weizenbaum repeatedly insisted, primitive. It scanned a user's typed input for keywords ranked by the active script; it applied decomposition rules to break the input into fragments; it applied reassembly rules to construct a response from those fragments. When no keyword was found, it fell back on stock phrases, "Please go on," "I see," "Tell me more", or on a memory structure that retrieved something the user had said earlier.¹¹

The design had an elegant evasion at its heart. Weizenbaum chose to make ELIZA's most famous script, DOCTOR, simulate a Rogerian psychotherapist because that style of therapy lets the therapist say almost nothing about the world while still appearing to listen, and a program that knows almost nothing about the world could do the same.¹² In Rogerian therapy the therapist reflects the patient's words back, asks open questions, and refrains from imposing an interpretive line. This is, mechanically speaking, almost exactly what a keyword-matching, phrase-reflecting program does anyway. The genius of the DOCTOR script was that it turned the program's ignorance into a feature. ELIZA did not need to know what "very unhappy these days" meant; it needed only to transform "I am very unhappy these days" into "How long have you been very unhappy these days?" and the illusion of understanding would, for many users, follow automatically.

Weizenbaum named the program after Eliza Doolittle, the character in Shaw's Pygmalion who is taught to pass as a lady of high society by being coached in the right forms of speech.¹³ The parallel was pointed. Like Eliza Doolittle, the program had learned, through a kind of elaborate coaching, to produce outputs that its interlocutors took as evidence of an inner life it did not have. The name was also, this essay's reading suggests, a mild joke at the expense of anyone who was taken in, a joke Weizenbaum would stop finding funny quite quickly.

Looking back in his 1976 book, Weizenbaum wrote that ELIZA had no built-in knowledge of any subject, and that the DOCTOR script demonstrated not that communication between man and machine was rich, but that it was "superficial."¹⁴ The 1966 paper's stated purpose was not to show that natural language understanding was achievable but to investigate the conditions under which people would attribute it to a machine that had not achieved it. Some subjects, he noted, had been "very hard to convince that ELIZA (with its present script) is not human", a sentence intended as evidence of human credulity, received almost everywhere as evidence of machine intelligence.¹⁵

What Happened Next

What happened next disturbed him profoundly. Users who knew they were talking to a program opened their hearts to it. His secretary asked for privacy. Psychiatrists saw clinical possibilities.¹⁶ Kenneth Colby, a former collaborator and friend of Weizenbaum who was then at Stanford, adapted the program's approach into a system called PARRY, designed to simulate a person with paranoid schizophrenia, and proposed that programs of this type could be deployed as therapeutic tools, scaled, automated, running in parallel on many terminals at once.¹⁷ Even Carl Sagan, no credulous enthusiast, imagined in 1975 networks of computer psychotherapy terminals where people could talk to "an attentive, tested and largely nondirective psychotherapist" for a few dollars a session.¹⁸

Weizenbaum's response to all of this was bafflement, and then something stronger. ELIZA, he said in 1984, "was instantly misunderstood as being basically the dawn of computerized psychiatry — which I detest." A few years later he described the idea of computer psychotherapy as "an obscene idea."¹⁹ The break with Colby was fast and bitter. In a small, parochial community, the rupture became personal. "The real break," Weizenbaum told Pamela McCorduck for her 1979 history of AI, "arose over the claim that ELIZA was of therapeutic significance."²⁰

What Weizenbaum found repugnant was not that a computer could be used in medical settings, he was not opposed to all clinical computing, but the specific substitution being proposed. To replace a human therapist with a machine was to replace a person who cares, or who must at least answer for whether they cared, with a system that has no access to caring whatsoever. The fact that users could not tell the difference did not mitigate this; it made it worse. The tendency to project understanding and feeling onto a system that has neither, formally defined by Douglas Hofstadter in Fluid Concepts and Creative Analogies (1995) as the "ELIZA effect", was, in Weizenbaum's view, a vulnerability in human beings, not a capability of machines.²¹ To exploit it therapeutically was to exploit a weakness in the very people who most needed genuine help.

Colby, for his part, dismissed Weizenbaum's objections as "pseudo-moralizing" and argued that withholding the opportunities the new technology offered would itself be immoral.²² The two men never reconciled.

The Distinction That Matters

The book Weizenbaum eventually published, Computer Power and Human Reason: From Judgment to Calculation, W. H. Freeman, 1976, is not a book against computers. This is important and often missed. Weizenbaum loved computers. He had spent his professional life building things with them. The MIT obituary described his technical contributions as genuine achievements and noted that he had entered computing in its earliest days with real enthusiasm.²³

Computer Power and Human Reason is, instead, a book against a specific confusion: between what computers can do and what computers ought to do. The central argument is that there are certain tasks that ought never to be delegated to machines regardless of whether machines can, technically, perform them, because those tasks require qualities that computers simply do not have: compassion, wisdom, moral judgment, and what he called the ability to confront genuine human problems in human terms.²⁴

Running beneath this is a philosophical distinction between deciding and choosing. Deciding, in Weizenbaum's account, is a computational activity: given a formal problem and an objective, a sufficiently powerful machine can find an optimal answer. Choosing is different. It is the product of judgment rather than calculation, and judgment is grounded in a life, in experience, in the particular history of the person making the choice. A computer, having no life in this sense, cannot make choices; it can only compute decisions. When we delegate to computers choices that ought to be choices, about punishment, about medical treatment, about the allocation of social resources, we do not remove human values from the process; we hide them. The values are still there, encoded in the design of the system.²⁵

Weizenbaum was explicit about which tasks he had in mind. A judge sentencing a criminal defendant exercises a form of care, attention to the particular person before the court, that cannot be reduced to a scoring formula. A psychotherapist does the same: the relationship itself is partly constitutive of the treatment. Replacing the judge or the therapist with a machine that computes a verdict or reflects a sentence does not merely automate the task; it eliminates the thing that made it what it was. This is more than distaste for modernity. It is a claim that some human practices are defined by the presence of a person who is genuinely at stake in the outcome.

Weizenbaum, who had been a child in Berlin in the 1930s and had watched a state perform enormous evil through the routine application of procedures, was not making an abstract point.²⁶

He was also alarmed, in the book, by what he called "compulsive programmers", the young men who inhabited the machine rooms at MIT and elsewhere, spending their lives in intimate, all-consuming relationship with computers to the exclusion of ordinary human contact. This is not a digression from the book's central argument but the psychological face of it: the hacker's obsessive relationship with calculation was, for Weizenbaum, another symptom of the same confusion between calculation and reason. In 1982, speaking to the Vancouver Sun, he said: "The computer is a particularly seductive instrument because it offers immediate feedback." And: "But the thing that binds you to do it is that whatever the computer is doing is under your control." It was a seduction he feared produced not power but dependency, not clarity but escape from the burden of being a person.²⁷

The Response from the Field

In November 1976, the Stanford Artificial Intelligence Laboratory published a pamphlet containing three simultaneous reviews of the book, by Bruce Buchanan, Joshua Lederberg, and John McCarthy. The other two were measured in tone; McCarthy's was not.²⁸ He titled his piece "An Unreasonable Book," and opened: "This moralistic and incoherent book uses computer science and technology as an illustration to support the view promoted by Lewis Mumford, Theodore Roszak, and Jacques Ellul, that science has led to an immoral view of man and the world."²⁹ The review ran to several pages, cataloguing Weizenbaum's polemical sins: unfair characterisations of colleagues, a "more-human-than-thou" attitude toward the AI researchers he criticised, and a habit of making extreme statements then qualifying them with contradictory ones.

Weizenbaum was not alone in attacking the pretensions of the field. Hubert Dreyfus had published What Computers Can't Do in 1972, arguing from phenomenological philosophy that AI was conceptually misconceived. But Dreyfus's claim was that machine intelligence was impossible in principle; Weizenbaum's was different and in some ways harder to dismiss, he was not saying that machines could not do these things, but that they should not, even if they could.

McCarthy also objected to what he called Weizenbaum's "biased, baseless, false, and seems motivated by political malice" suggestion that the Defense Department supported speech recognition research in order to eavesdrop on telephone calls.³⁰ On the specific question of Colby's proposal for computer psychotherapy, McCarthy was characteristically blunt: "Using computer programs as psychotherapists, as Colby proposed, would be moral if it would cure people. Unfortunately, computer science isn't up to it, and maybe the psychiatrists aren't either."³¹

The concession that Colby's therapeutic vision was premature was buried under the implication that Weizenbaum's objections were merely philosophical, feelings, not arguments. Lewis Mumford, who had written about technology and civilisation for decades, offered a contrasting assessment from the book's dust jacket, predicting it "will still be read half a century from now, as a landmark in the integration of scientific and humanistic thinking."³²

The book did cause disagreement with, and separation from, other members of the AI research community, a status Weizenbaum later said he had come to take pride in.³³ By the mid-1980s he was describing himself publicly as a heretic of technology.³⁴ In 1985, in New Age journal, he went further: "The dependence on computers is merely the most recent — and the most extreme — example of how man relies on technology in order to escape the burden of acting as an independent agent. It helps him avoid the task of giving meaning to his life, of deciding and pursuing what is truly valuable."³⁵

His political views deepened his estrangement from MIT, whose involvement in military research he opposed. He worried that what he called "listening machines" would eventually make voice surveillance easy and routine, a concern McCarthy dismissed as politically motivated, and which subsequent decades have rendered less easy to dismiss.³⁶ "I have pronounced heresy, and I am a heretic," he told the New York Times in 1977.³⁷

The Return

Weizenbaum retired from MIT in 1988, after a quarter-century on the faculty. In 1996, he moved back to Berlin, settling near his childhood neighbourhood in Mitte.³⁸ There he found an audience for his thinking and recognition as a public intellectual that had been harder to sustain in Cambridge, where he was known primarily as a dissenter from the mainstream of his own field. He died on 5 March 2008, aged 85, at Gröben in Brandenburg, and was buried at the Weißensee Jewish cemetery in Berlin.³⁹

In 2021, the researcher Jeff Shrager searched the MIT archives and found the original ELIZA source code in MAD-SLIP, with the DOCTOR script attached, in a folder labelled "Computer Conversations." The Weizenbaum estate gave permission to release the code under a Creative Commons public domain licence. In July 2026 Shrager and seven colleagues printed the rediscovered code, with scripts that had been missing for decades, in Inventing ELIZA (MIT Press). They argue that ELIZA was a more sophisticated system than had been documented, which sits awkwardly with Weizenbaum's own insistence that it was primitive.⁴⁰ In December 2024, engineers working with that code reconstructed the original ELIZA on an emulator of the MIT CTSS and IBM 7094, and demonstrated that it could reproduce almost exactly the sample conversations printed in Weizenbaum's 1966 paper.⁴¹ The code had been effectively lost for more than fifty years. Its recovery allowed scholars to compare what ELIZA had actually been with what it had become in reputation, and to confirm that the program's later fame as the founding chatbot had been shaped partly by the loss of access to the original artifact.

What He Got Right, and What He Got Wrong

McCarthy's counterattack was not baseless. The book is often vague where precision is needed. Weizenbaum's use of "obscenity" as a criterion for prohibiting research is, as McCarthy noted, loose enough to permit almost any interpretation. His portrait of the compulsive programmer, while recognisable in outline, slides from observation into caricature. His attribution to McCarthy of a quotation McCarthy denied ever saying, and which Weizenbaum himself acknowledged, after reviewing the tape of the Lighthill debate, was not on it, subsequently displacing the claim to another debate, gave McCarthy legitimate grounds for anger, and grounds to question the reliability of the book's other characterisations of colleagues.⁴²

What Weizenbaum failed to supply was a principle for drawing the line. Which decisions are really choices? Who gets to say? He did not have a full answer, and the vagueness left him exposed to the charge, McCarthy's charge, essentially, that he was simply opposed to things he found distasteful, dressed in the language of principle.

And yet. In the fifty years since Computer Power and Human Reason appeared, the systems Weizenbaum warned against have been built, in large numbers, in almost exactly the domains he named. Automated risk-scoring systems influence sentencing decisions in courts. Triage algorithms shape access to healthcare. Conversational systems of increasing sophistication, systems that make ELIZA look rudimentary, but that operate on the same basic principle of producing outputs that their interlocutors take as evidence of understanding, are now deployed as therapeutic companions, emotional support tools, and mental health interfaces, largely unregulated. The ELIZA effect operates at a scale Weizenbaum could not have imagined.

McCarthy's 1976 concession that computer psychotherapy was not yet technically feasible, "computer science isn't up to it", read then as a postponement rather than a refutation. What he was saying, in effect, was: object when the systems are inadequate in practice; once they are adequate, the objection disappears. Weizenbaum's point was that adequacy, in the clinical sense, is not and cannot be the right threshold, that there are things a machine ought not to do even when it can do them convincingly, because the convincingness is itself part of the problem.

Whether that argument holds is a philosophical question this essay cannot settle. What seems clear, and this is the essay's view, not the sources', is that the structure of the debate has not changed. On one side: if it works, use it. On the other: working is not enough.

Both sides skip the person who cannot pay for a human. For her, a machine that seems to listen may be better than nobody listening at all. And if human care becomes the luxury and machine care the cheap default, the argument stops being about what machines ought to do. It becomes about who gets the person and who gets the program.

"We've created a complex world which we have no control over anymore," he said in 2008, not long before his death. "No one understands them anymore, no one can understand them, because we have lost the information about their creation, the history of their creation, and that is a great danger for mankind."⁴³