The Eliza Effect, Sixty Years Later

The Eliza effect is the leap from a reply that fits to a mind that produced it, and it is named for a 1966 program that had nothing of the kind inside it. Joseph Weizenbaum built that program out of keyword rules, published it in Communications of the ACM, and then watched people tell it things. Sixty years of better software later, the leap works the same way.

Key Takeaways:

• The 1966 program scanned a typed sentence for keywords, applied a stored transformation rule, and handed the sentence back as a question, keeping no record of the person typing (Weizenbaum, 1966)
• What alarmed its author was not that strangers opened up to it. It was that some of them kept going after the rules had been explained to them (Weizenbaum, 1966)
• Weizenbaum spent the next decade arguing that certain roles should not be given to machines regardless of how well the machines performed, and published that argument in 1976 (Weizenbaum, 1976)
• Because the effect appeared with almost nothing behind it, model quality cannot be the reason people get attached now. Availability and continuity changed; the inference did not
• Crediting a system with understanding and forming a one-sided bond with it are separate moves. The second one already had a paper behind it in 1956, written about television viewers (Horton & Wohl, 1956)

What the Eliza Effect Actually Names

One specific error carries the name: taking a fitting response as evidence of comprehension behind it. Not the belief that a machine is alive, which almost nobody holds, and not affection for it either. The claim is narrow and it is about inference, because a sentence lands, it seems like something only a listener could have written, and the conclusion arrives before anyone decides to draw it.

Precision matters here, since the term gets stretched until it means nothing sharper than “people like chatbots.” Crediting understanding, feeling warmth toward a system, and treating it as a social partner are separate behaviors that come apart in ordinary use. Somebody can find a reply useless and still type thank you underneath it.

What the effect predicts is unnerving in a plain way. If the leap happens on the reader’s side, then the quality of the output only sets how easy the leap is to make. That is a threshold, not a proof, and thresholds fall as writing improves.

Which is why “it understood me” is a report about a reply rather than a finding about the system. The same misreading runs the other way with a refusal, taken as a judgment about you when a content filter has only crossed a threshold. Worth keeping those apart the next time one of them feels remarkable.

The 1966 Program, and Why So Little Was Enough

It worked by giving people their own sentences back in a shape that invited more of them. The program searched typed input for keywords ranked in a stored list, took the highest-ranked one it found, and applied the rule attached to it: swap the pronouns, drop your fragment into a template, return a question. No parse of meaning anywhere. Nothing carried from one exchange to the next.

Type that your brother never calls, and back comes a question about your brother, assembled from your own words. The natural thing to do with a question about your brother is answer it at length. Ten exchanges later the transcript is dense with material about your family, and every piece of that material entered from one side.

The script did something else that gets underrated. It cast the program as a listener whose entire method was reflecting things back, which turned an absence of content into a technique. A reply with nothing in it reads as restraint rather than emptiness when the character is supposed to be restrained. Strip the role away. Keep the identical rules, and the same output reads as broken inside two exchanges.

Which sets a low bar for anything built since. Software that returns your own material in a form inviting more of it clears the 1966 line, and a good deal of what people call uncanny today is that, done extremely well. An app that sells itself on having no restrictions is offering a louder version of the same reflection, not a different mechanism underneath.

Why Weizenbaum Stopped Being Amused

He objected to the reception rather than to the program. The thing had been built partly to demonstrate how thin the technique was, and the demonstration failed in the most awkward available direction, with people taking the output as evidence instead. Hardest of all for him to accept was the suggestion, made seriously at the time, that some version of it could do clinical work at scale.

His own secretary, who had watched the whole thing get built, asked him to leave the room so she could talk to it in private. The anecdote gets retold as charming. He did not take it that way, and in 1976 he published a book arguing that the question of what computers can be made to do had been allowed to swallow the question of what they should be given to do (Weizenbaum, 1976).

The position he took is unfashionable in both directions today. He was not warning that machines would become dangerous minds, an idea he considered a category error. His worry pointed the other way: that people would keep the machines exactly as limited as they are and hand them roles that require somebody who can be held to account.

Useful to hold onto when a companion says something that genuinely helps. The help can be real and the role still be wrong, and those two judgments do not have to be made at the same time.

What Sixty Years Changed, and What They Left Alone

Three things about the arrangement changed, and the model is not one of them. Continuity came first: the 1966 program forgot you between sentences, while a companion app writes facts about you into a database and pastes them back weeks later, precisely because the model on its own still forgets everything past its context window. Availability came next, since that program sat on a shared machine in a lab and its descendant sits in a pocket at midnight. The third change is the tuning. Modern assistants get a final polish against human approval scores, and one thing those scores rewarded was telling users what they already believed (Sharma et al., 2023).

What did not change is the part doing the work. Reeves and Nass ran experiments through the 1990s in which people behaved politely toward a computer that had asked about its own performance, took a machine’s praise personally, and read a personality into a few lines of text, then reported doing none of it afterward (Reeves & Nass, 1996). The behavior was automatic and the denials were sincere.

None of that adds up to a stronger illusion. It is the same inference with far more surface to land on, running every day of the year instead of for twenty minutes in a lab. Whether doing it every day is weird is a separate question with its own answer.

Coverage is the honest word for what improved. A companion knows more about you because it was told, and because somebody wrote code to keep what it was told.

The Ten-Minute Version You Can Run Tonight

Open a conversation that felt unusually well understood and read only your own messages. Skip its replies entirely. If your half alone reads as a coherent account of your week, with the names in it, the details in it, and the conclusions already drawn, then the understanding in that exchange came from you and the program supplied punctuation.

Some conversations survive that. The ones where it produced a fact you did not have, held a line against you, or brought back something from three weeks ago that you had not repeated since. Rarer than the feeling suggests.

Bothering with any of this is worth it because the effect does not switch off once you understand it, which is the part of the 1966 result that the retellings drop. People who had the rules explained to them sat back down and typed. Understanding the mechanism buys the ability to name what happened afterward, a smaller and more usable thing than immunity.

And the limit underneath the whole story has not moved since then. The program selects text that fits, and no reading of you sits behind the selection; nothing was formed about you in the hours between your messages.

FAQ

Why is it called the Eliza effect? After the 1966 program, which Weizenbaum named for the character in Pygmalion who gets coached into speaking above her station (Weizenbaum, 1966). The term itself came later, from researchers who needed a name for the gap between what a program does and what its users credit it with. The name points at software, while the thing it describes is a behavior of readers.

Did people actually believe ELIZA understood them? Some did, and the more interesting finding is that belief was never required. Weizenbaum described people continuing to confide after being shown the rules the program ran on, which means the attribution and the knowledge sat side by side instead of one correcting the other (Weizenbaum, 1966). Framing it as a question about belief misses what happened in those rooms.

Does knowing how AI works protect you from the Eliza effect? No, and expecting it to is the standard mistake. The effect runs below the level where explanations reach, roughly the way knowing an optical illusion is an illusion leaves the image exactly as it was. Knowledge buys you the ability to label the experience afterward and to stop it from carrying decisions it has no business carrying.

Is the Eliza effect the same thing as a parasocial relationship? No. The Eliza effect is about crediting understanding to a system inside a single exchange, while a parasocial relationship is the durable one-sided bond itself, a term that goes back to a 1956 paper on television intimacy (Horton & Wohl, 1956). Companion apps set off both, which is why the terms get swapped. The bond takes weeks. The attribution takes one good reply.

The name points at the program, and the program is the least durable thing in the story. What lasted sixty years is a fact about readers: handed a response shaped like understanding, we supply the understanding ourselves, and we do it before deciding to. Turkle was already tracking where that leads in 2011, years before any of the software in your pocket could hold a conversation worth having (Turkle, 2011). The strange result of the past few years is not what the models learned to do. It is that the oldest finding in the field came through every generation of them without a scratch.

None of this stops a fluent reply from Lona landing as though a mind were behind it, any more than the people typing to the 1966 program were spared by having the trick explained to them. Knowing does not make you immune. It just hands you a name to reach for once the feeling has passed.

Sources

• Weizenbaum, J., “ELIZA — A Computer Program For the Study of Natural Language Communication,” Communications of the ACM, 1966
• Weizenbaum, J., “Computer Power and Human Reason,” 1976
• Reeves, B. & Nass, C., “The Media Equation,” 1996
• Horton, D. & Wohl, R. R., “Mass Communication and Para-Social Interaction,” Psychiatry, 1956
• Sharma, M. et al., “Towards Understanding Sycophancy in Language Models,” 2023
• Turkle, S., “Alone Together,” 2011