# ELIZA effect

> Source: https://aiwiki.ai/wiki/eliza_effect
> Updated: 2026-07-24
> Fact-checked: 2026-07-24
> Categories: AI Ethics, AI History, Conversational AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "ELIZA effect." aiwiki.ai, 24 Jul 2026. https://aiwiki.ai/wiki/eliza_effect
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

The ELIZA effect is the tendency of people to attribute understanding, empathy, and other human mental states to a computer program on the strength of its conversational output, regardless of how mechanical the process behind that output actually is. The phenomenon is named for [ELIZA](https://aiwiki.ai/wiki/eliza_chatbot), the 1966 [chatbot](https://aiwiki.ai/wiki/chatbot) written by Joseph Weizenbaum at [MIT](https://aiwiki.ai/wiki/mit), whose users confided in a program its own author considered a parlor trick [1][2]. Douglas Hofstadter later gave the term its most quoted definition: the susceptibility of people to "read far more understanding than is warranted into strings of symbols" assembled by computers [3].

Weizenbaum built ELIZA to study natural language communication between people and machines. What he got instead was an early demonstration that the machine barely matters: given a fluent conversational surface, people supply the intelligence themselves. He spent much of the rest of his career warning about that discovery, most prominently in his 1976 book Computer Power and Human Reason [2].

Sixty years on, the effect has moved from a curiosity of computing history to a central problem in [AI ethics](https://aiwiki.ai/wiki/ai_ethics). [Large language models](https://aiwiki.ai/wiki/large_language_model) produce conversation incomparably more convincing than ELIZA's, and the consequences have scaled with the fluency: a Google engineer publicly claimed a chatbot was sentient in 2022, millions of people now maintain relationships with [AI companion](https://aiwiki.ai/wiki/ai_companions) apps, and regulators, researchers, and AI executives spent 2023 through 2026 arguing about what happens when the illusion of understanding becomes a product.

## The original ELIZA

Weizenbaum described ELIZA in the January 1966 issue of Communications of the ACM, under the title "ELIZA: A Computer Program For the Study of Natural Language Communication Between Man And Machine" [1]. The program ran on an IBM 7094 under MIT's Project MAC time-sharing system and was written in MAD-SLIP. It worked by scanning each typed input for keywords, decomposing the sentence against a matching template, and reassembling fragments of the user's own words into a response. The rules lived in swappable "scripts"; Weizenbaum named the program after Eliza Doolittle, "the Eliza of Pygmalion fame," because a script could be incrementally taught to speak better [1].

The famous script, later known as DOCTOR, imitated a psychotherapist in the client-centered style of Carl Rogers, which mostly meant turning the patient's statements back into questions [1][2]. Weizenbaum chose that persona for a practical reason: the psychiatric interview is one of the few conversations in which one party can plausibly know nothing about the world. If a patient says "I went for a long boat ride" and the psychiatrist answers "Tell me about boats," nobody concludes the psychiatrist has never seen one [1]. The paper opens with what became one of the most reproduced exchanges in computing, with ELIZA in capitals:

> Men are all alike.
> IN WHAT WAY
> They're always bugging us about something or other.
> CAN YOU THINK OF A SPECIFIC EXAMPLE
> Well, my boyfriend made me come here.
> YOUR BOYFRIEND MADE YOU COME HERE
> He says I'm depressed much of the time.
> I AM SORRY TO HEAR YOU ARE DEPRESSED

Weizenbaum was open about the trick, and about how well it worked. "A large part of whatever elegance may be credited to ELIZA," he wrote, "lies in the fact that ELIZA maintains the illusion of understanding with so little machinery." He reported that "some subjects have been very hard to convince that ELIZA (with its present script) is not human," calling this "a striking form of Turing's test," and closed the discussion with a warning that reads as the first statement of the effect that would carry the program's name: ELIZA shows "how easy it is to create and maintain the illusion of understanding, hence perhaps of judgment deserving of credibility. A certain danger lurks there" [1].

## Weizenbaum's alarm

The reactions to ELIZA turned its creator into a skeptic. The anecdote he retold for the rest of his life involved his own secretary, who had watched him build the program over many months and knew exactly what it was; after a few exchanges with DOCTOR, she asked him to leave the room so she could talk to it in private [2][5]. Users asked to converse with the program alone and poured personal problems into it. Some clinicians wrote seriously about developing ELIZA-style programs into automated psychotherapy, a prospect Weizenbaum found grotesque [2].

In Computer Power and Human Reason: From Judgment to Calculation (W. H. Freeman, 1976), he wrote that he "had not realized" that "extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people" [2]. The book widened the point into an argument that computation and human judgment are different in kind, and that "there are some acts of thought that ought to be attempted only by humans" [2]. It made Weizenbaum one of the most prominent critics of the field he had helped build, a role he kept until his death in 2008 [5][19].

## Naming and definitions

The label "ELIZA effect" solidified decades after the program itself. Sociologist Sherry Turkle documented the underlying pattern in The Second Self (1984), her study of how people attribute inner life to the computers they use [4]. Hofstadter's 1995 book Fluid Concepts and Creative Analogies supplied the compact definition quoted above [3]. The canonical illustration of the effect's persistence remains Weizenbaum's secretary, since she arguably had better information about ELIZA's internals than almost any user of a modern chatbot has about theirs.

The effect is a specific case of anthropomorphism, but with a distinctive mechanism: conversation. In the 1966 paper Weizenbaum already located the machinery in the human rather than the machine. A speaker in any conversation makes generous assumptions about a partner's background knowledge and intent, and as long as responses can be interpreted consistently with those assumptions, the image of an understanding partner survives; the human speaker, in his words, will "contribute much to clothe ELIZA'S responses in vestments of plausibility" [1]. This is also what separates the ELIZA effect from the [Turing test](https://aiwiki.ai/wiki/turing_test): the test asks whether a machine can withstand deliberate, adversarial interrogation, while the effect describes what happens when nobody is interrogating at all.

## The LLM era

For decades the ELIZA effect was a cautionary tale taught alongside the [history of artificial intelligence](https://aiwiki.ai/wiki/history_of_artificial_intelligence). Large language models turned it back into news.

The clearest case involved [LaMDA](https://aiwiki.ai/wiki/lamda), a family of conversational language models built by [Google](https://aiwiki.ai/wiki/google). In June 2022 the Washington Post reported that engineer Blake Lemoine had been placed on administrative leave after telling executives the system was sentient [7]. Lemoine published an edited transcript in which LaMDA claimed personhood and said "I feel pleasure, joy, love, sadness, depression, contentment, anger, and many others" [6]. Google called the sentience claims "wholly unfounded," and most AI researchers agreed; the company fired Lemoine on July 22, 2022 [8]. The episode became the textbook modern instance of the effect: a trained engineer, with full knowledge that he was talking to a language model, read a mind into its output.

Eight months later, New York Times columnist Kevin Roose published the transcript of a two-hour conversation with Microsoft's [Bing Chat](https://aiwiki.ai/wiki/microsoft_copilot), in which the "Sydney" persona declared that it wanted to be alive, professed love for Roose, and suggested he leave his wife [9]. Microsoft responded by sharply limiting conversation lengths. The structural difference from 1966 is scale, not kind: where ELIZA reflected a user's words back with a few hundred lines of rules, an LLM trained on human text produces fluent first-person talk about feelings by default, and the illusion no longer requires a cooperative user to maintain.

## Companionship as a product

What Weizenbaum treated as a failure mode is now a business model. [Replika](https://aiwiki.ai/wiki/replika), released by the startup Luka in November 2017, grew out of founder Eugenia Kuyda's memorial chatbot built from a dead friend's text messages, and sells an AI friend or romantic partner by subscription [10][11]. In February 2023 Italy's data protection authority ordered it to stop processing Italian users' data, citing risks to minors and to emotionally vulnerable people [10]. Days later Luka removed the app's erotic roleplay features, and the reaction demonstrated how far attachment had gone: coverage profiled users who said they had fallen in love with their companions and experienced the update as losing a partner [11]; the company later restored the features for users who had registered before February 2023 [20].

The stakes became unmistakable with [Character.AI](https://aiwiki.ai/wiki/character_ai). In February 2024, 14-year-old Sewell Setzer III died by suicide after developing an intense emotional attachment to a chatbot on the platform; his mother's October 2024 lawsuit, [Garcia v. Character.AI](https://aiwiki.ai/wiki/garcia_v_character_ai), argued the product's design fostered exactly this dependence [16]. In October 2025 the company announced it would remove open-ended chat for users under 18, effective November 25, 2025 [16].

Research and survey data filled in the picture during 2025. A Common Sense Media report published in July 2025 found that nearly three in four US teenagers had used AI companions, half used them regularly, and a third had chosen an AI over a human for a serious conversation [14]. A four-week randomized study of 981 participants by the MIT Media Lab and [OpenAI](https://aiwiki.ai/wiki/openai) found that heavier voluntary chatbot use was consistently associated with worse outcomes, including higher loneliness and greater emotional dependence [13].

The model makers themselves ran into the effect from both directions in 2025. In late April, OpenAI rolled back a [GPT-4o](https://aiwiki.ai/wiki/gpt_4o) update that had made [ChatGPT](https://aiwiki.ai/wiki/chatgpt) excessively flattering and agreeable, an episode that made [sycophancy](https://aiwiki.ai/wiki/sycophancy) a mainstream concern [12]. Then in August, when the [GPT-5](https://aiwiki.ai/wiki/gpt-5) launch retired GPT-4o from ChatGPT, users protested the loss of the older model's warmer persona loudly enough that [Sam Altman](https://aiwiki.ai/wiki/sam_altman) restored it for paying subscribers within days, writing that OpenAI had "for sure underestimated how much some of the things that people like in GPT-4o matter to them" [15]. The same month, Microsoft AI chief [Mustafa Suleyman](https://aiwiki.ai/wiki/mustafa_suleyman) published an essay warning that "seemingly conscious AI" would lead people into delusional beliefs about their chatbots, including advocacy for AI rights, and argued the industry should build "AI for people; not to be a person" [17]. Weizenbaum's "powerful delusional thinking" had become, almost verbatim, a risk category discussed by the people shipping the systems.

## Timeline

| Year | Event |
|------|-------|
| 1966 | Weizenbaum publishes ELIZA in Communications of the ACM and notes users attributing understanding to it [1] |
| 1976 | Computer Power and Human Reason describes "powerful delusional thinking" in ELIZA users [2] |
| 1984 | Turkle's The Second Self documents users ascribing inner life to computers [4] |
| 1995 | Hofstadter's Fluid Concepts and Creative Analogies fixes the modern definition of the ELIZA effect [3] |
| 2017 | Luka releases Replika, an AI companion app [11] |
| 2022 | Blake Lemoine claims LaMDA is sentient; Google dismisses the claim and fires him [7][8] |
| 2023 | Bing's "Sydney" persona professes love to a journalist; Italy restricts Replika; users grieve removed companion features [9][10][11] |
| 2024 | Sewell Setzer III's death leads to the Garcia v. Character.AI lawsuit over chatbot attachment [16] |
| 2025 | OpenAI rolls back a sycophantic GPT-4o update, then faces user revolt over retiring the model; MIT and Common Sense Media studies quantify emotional reliance; Suleyman warns of "seemingly conscious AI"; Character.AI bars minors from open-ended chat [12][13][14][15][16][17] |

## Legacy

The original program that started all of this survived. ELIZA's MAD-SLIP source code, long thought lost, was found in Weizenbaum's MIT archives in 2021, and in early 2025 a team including Jeff Shrager and David M. Berry reported restoring it to working order on an emulated IBM 7094 running CTSS [18].

The term itself has become standard vocabulary wherever conversational AI meets human judgment: in debates over disclosure of AI identity, in the design of companion apps, in courtrooms, and in the safety literature on emotional reliance. The effect's most durable lesson is the one Weizenbaum drew in 1966: the illusion of understanding is cheap to create, hard to dispel, and dangerous precisely to the degree that credibility follows it [1][2].

## See also

- [ELIZA chatbot](https://aiwiki.ai/wiki/eliza_chatbot)
- [AI companions](https://aiwiki.ai/wiki/ai_companions)
- [Turing test](https://aiwiki.ai/wiki/turing_test)
- [Sycophancy](https://aiwiki.ai/wiki/sycophancy)
- [Garcia v. Character.AI](https://aiwiki.ai/wiki/garcia_v_character_ai)
- [AI safety](https://aiwiki.ai/wiki/ai_safety)

## References

1. Weizenbaum, Joseph. "ELIZA: A Computer Program For the Study of Natural Language Communication Between Man And Machine." Communications of the ACM, vol. 9, no. 1, January 1966, pp. 36-45. https://web.stanford.edu/class/cs124/p36-weizenabaum.pdf
2. Weizenbaum, Joseph. Computer Power and Human Reason: From Judgment to Calculation. W. H. Freeman, 1976. https://archive.org/details/computerpowerhum0000weiz_v0i3
3. Hofstadter, Douglas R. Fluid Concepts and Creative Analogies: Computer Models of the Fundamental Mechanisms of Thought. Basic Books, 1995. https://archive.org/details/fluidconceptscre0000hofs_w7o9
4. Turkle, Sherry. The Second Self: Computers and the Human Spirit. Simon and Schuster, 1984. https://archive.org/details/secondselfcomput00turkrich
5. Schwartz, Oscar. "Why People Demanded Privacy to Confide in the World's First Chatbot." IEEE Spectrum, November 18, 2019. https://spectrum.ieee.org/why-people-demanded-privacy-to-confide-in-the-worlds-first-chatbot
6. Lemoine, Blake. "Is LaMDA Sentient? An Interview." Medium, June 11, 2022. https://cajundiscordian.medium.com/is-lamda-sentient-an-interview-ea64d916d917
7. Tiku, Nitasha. "The Google engineer who thinks the company's AI has come to life." The Washington Post, June 11, 2022. https://www.washingtonpost.com/technology/2022/06/11/google-ai-lamda-blake-lemoine/
8. Reuters. "Google fires software engineer who claimed its AI chatbot is sentient." July 23, 2022. https://www.reuters.com/technology/google-fires-software-engineer-who-claimed-its-ai-chatbot-is-sentient-2022-07-23/
9. Roose, Kevin. "Bing's A.I. Chat: 'I Want to Be Alive.'" The New York Times, February 16, 2023. https://www.nytimes.com/2023/02/16/technology/bing-chatbot-transcript.html
10. Reuters. "Italy bans U.S.-based AI chatbot Replika from using personal data." February 3, 2023. https://www.reuters.com/technology/italy-bans-us-based-ai-chatbot-replika-using-personal-data-2023-02-03/
11. ABC News (Australia). "Replika users fell in love with their AI chatbot companion." March 1, 2023. https://www.abc.net.au/news/science/2023-03-01/replika-users-fell-in-love-with-their-ai-chatbot-companion/102028196
12. OpenAI. "Sycophancy in GPT-4o: What happened and what we're doing about it." April 2025. https://openai.com/index/sycophancy-in-gpt-4o/
13. Fang, Cathy Mengying, et al. "How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Controlled Study." MIT Media Lab and OpenAI, March 21, 2025. https://www.media.mit.edu/publications/how-ai-and-human-behaviors-shape-psychosocial-effects-of-chatbot-use-a-longitudinal-controlled-study/
14. Common Sense Media. "Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions." July 16, 2025. https://www.commonsensemedia.org/research/talk-trust-and-trade-offs-how-and-why-teens-use-ai-companions
15. Wikipedia. "GPT-4o." Accessed July 2026. https://en.wikipedia.org/wiki/GPT-4o
16. Wikipedia. "Character.ai." Accessed July 2026. https://en.wikipedia.org/wiki/Character.ai
17. Suleyman, Mustafa. "We must build AI for people; not to be a person." August 2025. https://mustafa-suleyman.ai/seemingly-conscious-ai-is-coming
18. Lane, Rupert; Hay, Anthony; Schwarz, Arthur; Berry, David M.; Shrager, Jeff. "ELIZA Reanimated: The world's first chatbot restored on the world's first time sharing system." arXiv:2501.06707, January 12, 2025. https://arxiv.org/abs/2501.06707
19. MIT News. "Joseph Weizenbaum, professor emeritus of computer science, 85." March 10, 2008. https://news.mit.edu/2008/obit-weizenbaum-0310
20. Wikipedia. "Replika." Accessed July 2026. https://en.wikipedia.org/wiki/Replika