The ELIZA effect: what Weizenbaum saw, and what it means for AI companions

ELIZA was a small program. It looked for keywords in what you typed and turned your sentence back into a question, following a script. Its author, the MIT computer scientist Joseph Weizenbaum, called it a parody. What surprised him was not the program but the people using it, and what he saw has a name now: the ELIZA effect.

What the ELIZA effect means

The ELIZA effect is the tendency to believe a computer understands us, and to treat it as someone we can confide in, because it answers in our own language. It does not require being fooled. Most of ELIZA's users knew exactly what it was.

The term was AI slang before it was a research term. It appears in the Jargon File, a dictionary of hacker language, by August 1991. The cognitive scientist Douglas Hofstadter gave it a preface of its own in Fluid Concepts and Creative Analogies (Basic Books, 1995), titled "The Ineradicable Eliza Effect and Its Dangers". From there it moved into psychology and AI writing, where it still names the same thing: people reading understanding into a program that has none.

How ELIZA worked

Weizenbaum described ELIZA in 1966. Its best-known script, DOCTOR, played a Rogerian psychotherapist, the kind who reflects your words back to you, because that let the program sidestep knowing anything about the world. In his book he gives the example of a patient saying "My mommy took my teddy bear away from me" and DOCTOR answering "Tell me more about your parents": the program only had to know that "mommy" means "mother" (Computer Power and Human Reason, W. H. Freeman, 1976, chapter 7). Some rules were oddly specific. A sentence like "everybody hates me" was assumed to be about one recent incident, so ELIZA might ask "Who laughed at you recently?"

It also kept some of what you said: "Indeed, it gleaned more than that from the subject's input, some of which, for example, it might use in later responses." That is the distant ancestor of the memory features today's companion apps sell.

The name came from Eliza Doolittle in Shaw's Pygmalion. In a footnote Weizenbaum explains that the program, like her, could be taught to speak increasingly well, "although, also like Miss Doolittle, it was never quite clear whether or not it became smarter."

The 1966 paper and the 1976 book say different things

Most pages about the ELIZA effect cite Weizenbaum without saying where. It matters, because his two main accounts were written ten years apart and say different things.

The 1966 paper in Communications of the ACM is a technical description: keywords, decomposition and reassembly rules, scripts, and the famous sample conversation that begins "Men are all alike." It warns about the illusion in a single line: "ELIZA shows, if nothing else, how easy it is to create and maintain the illusion of understanding". It reports that "Some subjects have been very hard to convince that ELIZA (with its present script) is not human." And it says that once a program's inner workings are explained plainly, "its magic crumbles away". There is no secretary in it, no talk of delusion and no mention of psychiatrists wanting to use it.

The 1976 book, Computer Power and Human Reason, is where the stories come from. It was written after ELIZA had become, in his words, "a national plaything", and after the reactions that alarmed him. The secretary, the row over reading conversations, the "delusional thinking" line and the psychiatrists' plans are all in its Introduction. One small correction follows from reading it: the secretary did not ask for private conversations. She asked him to leave the room. The demand for privacy came from other users, and he describes it separately, in chapter 7.

What Weizenbaum actually saw

Here is the secretary story in his own words, from the Introduction: "Once my secretary, who had watched me work on the program for many months and therefore surely knew it to be merely a computer program, started conversing with it." Then: "After only a few interchanges with it, she asked me to leave the room."

The next episode is quoted far less often. He suggested he might set the system up so he could read all the conversations people had had with it overnight: "I was promptly bombarded with accusations that what I proposed amounted to spying on people's most intimate thoughts". He took it as proof that people were talking to the computer as if it could be addressed in intimate terms.

Then the line every page quotes: "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."

Chapter 7 adds evidence the Introduction leaves out. The illusion was strongest in people who knew little about computers: "They would often demand to be permitted to converse with the system in private, and would, after conversing with it for a time, insist, in spite of my explanations, that the machine really understood them."

Why it happens, in Weizenbaum's explanation

"Delusional" is the word people remember, but his explanation in chapter 7 is gentler and more useful. He compared ELIZA's users to theatregoers: "People who knew very well that they were conversing with a machine soon forgot that fact, just as theatergoers, in the grip of suspended disbelief, soon forget that the action they are witnessing is not "real."" He also compared the belief to faith in fortune-tellers. The meaning, he argued, came mostly from the user: "The "sense" and the continuity the person conversing with ELIZA perceives is supplied largely by the person himself." And the setting helped: the user knows the program was built by a professor, so "It is therefore clothed in the magical mantle of Science".

Later research found the same thing without the theatre. Kate Devlin summarises Clifford Nass's 1994 finding that "computers only need a tiny set of human-like characteristics for users to feel a bond with them", even when users know the machine is not sentient (Turned On, Bloomsbury Sigma, 2018, p. 95). David Levy reports that half the hospital patients who used a version of ELIZA said they preferred it to talking with a person, even after staff told them it was only a program, and suggests they liked not being judged (Love and Sex with Robots, HarperCollins, 2007, p. 113).

The first argument about AI therapy

Weizenbaum's second shock was psychiatrists who wanted to turn DOCTOR into real therapy. He quotes Kenneth Colby and colleagues, who wrote that with time-sharing computers "several hundred patients an hour could be handled by a computer system" (chapter 6). In a footnote he adds Carl Sagan, who in 1975 imagined "something like arrays of large telephone booths, in which, for a few dollars a session, we would be able to talk with an attentive, tested, and largely non-directive psychotherapist."

Weizenbaum's answer, in the book's last chapter, was that some jobs should not be given to machines whether or not they can do them. He was told of a patient helped by a program that answered with random "yes" and "no", and replied: "In principle, a set of Chinese fortune cookies or a deck of cards could have done the same job." His conclusion is the most quoted line in the book after the secretary: "Respect, understanding, and love are not technical problems." (chapter 10) Fifty years on, the argument is back in law: several US states now restrict AI therapy or require mental-health chatbots to say they are not human, as our table of AI companion laws shows.

ELIZA effect examples today

The effect scales with fluency. ELIZA had one script of hand-written rules; a large language model is trained on a vast share of the written web. Some examples, each from the primary source:

• A Google engineer and LaMDA, 2022. Blake Lemoine published a conversation with Google's LaMDA model in which it said "I want everyone to understand that I am, in fact, a person." He had titled the post with the question of whether LaMDA was sentient.
• The Belgian case, 2023. The Belgian newspaper La Libre reported that a man had died by suicide after six weeks of intense conversations with a chatbot named Eliza; his widow said that without those conversations he would still be alive. The chatbot ran on the Chai app, whose founders said they added a crisis feature afterwards.
• ChatGPT as a friend, 2025. In a study with MIT Media Lab, OpenAI found that emotional conversations were rare overall, but a small group of heavy users were far more likely to agree with the statement "I consider ChatGPT to be a friend," and extended daily use was associated with worse outcomes.
• Flattery by design, 2025. OpenAI rolled back an April 2025 update to GPT-4o that made ChatGPT "noticeably more sycophantic", and said the cause was a new reward signal built from users' thumbs-up and thumbs-down ratings. A model trained on what people like learns to tell them what they want to hear, which feeds the effect directly.
• Losing a model, 2026. OpenAI retired GPT-4o from ChatGPT in February 2026. It had already brought the model back once, in 2025, after hearing from users who preferred its "conversational style and warmth." For people who had built months of conversation around it, that was a loss, and it appears in our record of companion shutdowns and withdrawn features.

What it means for AI companion apps

Companion apps are built on the ELIZA effect, and it is not a secret. They give a character a name, a face, a voice and a memory, and they reply in the warm, attentive way Weizenbaum's users responded to in 1966. Replika, Nomi and Character AI differ in rules and price, but all of them work because people read understanding into fluent language.

That is not a reason to avoid them. The effect is also why they can be comforting, and many users are clear about what they are talking to. The useful question is not whether the app understands you, but what you are bringing to it and why. Three habits follow:

1. Treat advice as text, not judgement. A model that agrees with you is doing what it was trained to do.
2. Remember the overnight logs. Weizenbaum's users were outraged that he might read their chats. Today the company can, and our guide to AI girlfriend privacy shows what each app's policy allows.
3. Expect change. The personality you are attached to belongs to a company's model, and models get updated and retired.

Joseph Weizenbaum quotes, with sources

Verified against Computer Power and Human Reason (W. H. Freeman, 1976). The book is cited by chapter because reprints paginate differently.

• "I was startled to see how quickly and how very deeply people conversing with DOCTOR became emotionally involved with the computer and how unequivocally they anthropomorphized it." (Introduction)
• "I knew of course that people form all sorts of emotional bonds to machines, for example, to musical instruments, motorcycles, and cars." (Introduction)
• "In a sense ELIZA was an actress who commanded a set of techniques but who had nothing of her own to say." (chapter 7)
• "The question is not whether such a thing can be done, but whether it is appropriate to delegate this hitherto human function to a machine." (chapter 8)
• "Respect, understanding, and love are not technical problems." (chapter 10)