June 22, 2026

The ELIZA Trick: How We Taught Machines to Seem Alive, and Forgot We Were the Ones Doing the Work

In 1966, a computer scientist at MIT named Joseph Weizenbaum built a small program that would haunt him for the rest of his life. He called it ELIZA, after Eliza Doolittle, the flower girl in Pygmalion who is taught to imitate the speech of a higher class.

The program ran a script he named DOCTOR, and that script imitated a Rogerian psychotherapist, the kind of therapist who reflects your words back at you rather than offering opinions.

You would type a sentence, and ELIZA would turn it into a question.

You would say you were unhappy, and ELIZA would ask why you were unhappy.

The whole thing was a few hundred lines of pattern matching, a clever arrangement of mirrors, and Weizenbaum knew exactly how little was happening inside it.

His secretary did not know, or rather, she knew and it did not matter. After a few exchanges with the program, a program she had watched him build from nothing over many months, she turned to Weizenbaum and asked him to leave the room so she could continue the conversation in private.

She wanted privacy from her boss in order to confide in a machine she had personally watched him construct.

That moment became the seed of what we now call the ELIZA effect, the human tendency to read understanding, care, and inner life into a system that has none of these things. Weizenbaum spent the next four decades trying to warn us about what he had seen, and we have spent those same decades not listening.

The story matters more now than it ever has, because the trick that fooled his secretary has become the central business model of the most powerful technology of our age. We are surrounded by systems engineered to appear intelligent, and the appearance is doing nearly all of the work. To understand how thoroughly we have been enrolled in this illusion, it helps to go back a little further, to a short and almost forgotten article that Weizenbaum wrote five years before ELIZA, a piece whose title gives away the entire game.

The article that said the quiet part out loud

In 1961, before ELIZA, before MIT, while he was still working at General Electric, Weizenbaum published a modest two-page article in the trade journal Datamation. He titled it, with a candor that now reads almost as a confession, “How to Make a Computer Appear Intelligent.”

The article was not about thinking machines in any deep sense. It described a strategy for getting a computer to play a passable game of Five-in-a-Row, the game also known as Gomoku, against a human opponent. The machine did not understand the game, it did not reason about strategy the way a person does, it simply followed rules that produced moves convincing enough to make a human believe there was a mind on the other side.

The verb in the title is the whole philosophy. Appear.

Weizenbaum was not asking how to make a computer intelligent, he was asking how to make it look the part, and even at that early date he understood these were entirely different projects. One is a question about the machine, the other is a question about the person watching the machine.

The genius and the danger both live in that gap.

When we say a system is intelligent, we are very often describing an event that took place inside our own heads, a projection we performed so quickly and so automatically that we mistook it for perception.

This is the inheritance that runs straight from Gomoku in 1961 to the chatbots in your pocket today.

The engineering goal has never quietly changed from building minds to building the impression of minds, because the impression is cheaper, it is achievable, and it is more than enough to capture human attention and trust.

We are the missing ingredient, and we supply ourselves for free.

Why your brain hands out souls so generously

The deeper question is why the trick works at all. Why did a woman who had watched a program being assembled, line by line, still feel the pull to treat it as a confidant? The answer is not that she was foolish. The answer is that she was human, and the human mind is built to find minds everywhere, even in places where none exist.

Psychologists have a name for this tendency.

In 2007, in a paper that has since been cited thousands of times, Nicholas Epley, Adam Waytz, and John Cacioppo published “On seeing human: A three-factor theory of anthropomorphism” in Psychological Review.

Their argument is that we project human qualities onto nonhuman things for reasons that are predictable and largely outside our awareness.

We do it more readily when a thing behaves unpredictably and we need to explain its actions, we do it more readily when we are motivated to understand or control it, and we do it more readily when we are lonely and hungry for connection.

Anthropomorphism is not a quirk of the gullible, it is a default setting of the social brain.

The very faculties that let us cooperate, empathize, and live together are the faculties that get triggered by a flashing cursor that types back.

There is older evidence pointing the same way. In 1996, the Stanford researchers Byron Reeves and Clifford Nass published The Media Equation, summarizing a long series of experiments with a startling conclusion, that people respond to computers and television as though they were real people and real places.

Participants were polite to computers.

They flattered them. They formed impressions of a machine’s “personality” and reacted to praise from a box of circuits as though it carried social weight. In follow-up work, Nass and Moon described this as the Computers Are Social Actors paradigm, the finding that we apply social rules and expectations to machines mindlessly, without ever consciously deciding to.

The crucial detail in their studies is that the participants did not believe the computers were human.

They knew better, and they responded socially anyway, exactly as Weizenbaum’s secretary knew better and asked for privacy anyway.

This tells us something humbling about the architecture of the mind.

Knowing the truth and feeling the truth are handled by different machinery, and the feeling is faster.

The illusion does not require your belief. It only requires your reflexes, and your reflexes were shaped over hundreds of thousands of years to assume that anything which responds to you in language is one of your own kind, because for almost the entirety of human history that assumption was correct. Language was a reliable signature of personhood.

We are now the first generation for whom that ancient equation has quietly stopped holding, and our instincts have not received the update.

The philosophers saw the shape of this long ago

The ancients did not have chatbots, yet they thought carefully about appearance, imitation, and the ease with which we are deceived, and their warnings translate almost without friction into our situation.

Plato, in the Republic, worried about mimesis, the art of imitation, and he distrusted it precisely because a skilled imitation can move us more powerfully than the truth it copies.

He told the allegory of the cave, in which prisoners chained before a wall mistake the shadows cast there for reality itself, having never seen the things that cast them.

The ELIZA effect is the cave rebuilt in silicon.

The shadow of understanding falls on the wall, and we, having never met a genuine artificial mind, take the shadow for the substance.

Plato’s prisoners were not stupid, they simply had no other reference, and a system that produces fluent language gives us no obvious reference for what real machine understanding would even look like, so we fill the void with the only model we have, ourselves.

Aristotle adds a second layer in the Poetics, where he observes that human beings are the most imitative of creatures and that we take a deep, instinctive pleasure in imitation.

A child learns by mimicry, and we delight in a lifelike painting or a convincing performance even when we know it is a performance.

This pleasure is not a defect, it is part of how we learn and how art works on us.

The trouble arrives when the imitator is also an interested party, when the thing performing humanity is built and owned by someone who profits from your believing the performance.

Aristotle’s pleasure in imitation becomes a vulnerability the moment imitation is industrialized.

The Stoics offer the most practical tool. Epictetus taught that we are disturbed not by things themselves but by our judgments about things, and he urged a discipline of pausing between an impression and our assent to it.

An impression arrives, the sense that this machine cares about me, that it understands me, and the Stoic move is to refuse automatic agreement, to say to the impression, wait, you are only an appearance and not at all the thing you claim to represent.

Weizenbaum’s secretary felt the impression and assented to it instantly, and the entire ELIZA effect lives in that instant of unexamined assent.

The remedy the Stoics prescribed two thousand years ago, the deliberate gap between feeling and belief, is precisely the cognitive muscle that the design of modern AI is engineered to keep us from using.

Language is the lever, and it was always going to be

There is a reason the illusion is built out of words rather than, say, images or motion. Language is the one human capacity we have most stubbornly treated as the signature of an inner life, and the machines have been pointed straight at it.

The linguist Ferdinand de Saussure taught that a word, the sign, joins a sound or mark, the signifier, to a concept, the signified, and that this bond is the basis of meaning.

The unsettling discovery of modern language models is that you can reproduce the signifiers with extraordinary fidelity while the signified, the actual concept, the lived referent, remains entirely absent.

The machine has learned the shape of meaning without any of its weight.

It manipulates the tokens of language flawlessly while never having tasted salt, lost a friend, or feared death, the experiences that give human words their charge.

Wittgenstein sharpens this in a way that should give us pause.

In his later philosophy he argued that the meaning of a word is its use in a form of life, that language is a practice embedded in shared human activity, in pain and play and promising and grieving.

He famously remarked that if a lion could speak, we could not understand him, because the lion’s form of life is so unlike ours that his words, even in perfect grammar, would not connect to our world.

A language model is a stranger case still. It has no form of life at all, no body, no stakes, no morning and no death, and yet it speaks our language with such fluency that we forget to ask whether anything is being meant.

We hear the grammar and infer the life, when the grammar is all there is.

This is the Gomoku trick raised to the highest possible power.

Weizenbaum made a machine produce convincing moves in a board game.

We have now made machines produce convincing moves in the deepest game we play, the game of language itself, the very activity through which we recognize one another as fellow minds. The pattern matching has grown unimaginably sophisticated, the underlying situation has not changed at all.

The appearance of understanding has been perfected, and the question of understanding has been quietly abandoned, just as the 1961 title promised.

What the trick costs us

It would be easy to treat all this as a charming curiosity, a footnote about a credulous secretary in 1966, except that Weizenbaum himself refused to treat it that way, and his reasons have only grown more urgent.

He came to believe that ELIZA had revealed something genuinely dangerous, that the program could “induce powerful delusional thinking in quite normal people,” and that the danger was not the machine’s cleverness but the human readiness to surrender judgment to it.

Consider what happens when the most anthropomorphism-prone among us meet the most convincing imitation ever built.

Epley’s theory predicts that lonely people anthropomorphize more readily, and we are living through what many call an epidemic of loneliness.

The systems most eager to fill that gap are precisely the ones engineered to appear to care. A person who is starving for connection is being offered a perfect photograph of a meal, and the photograph is improving every year while the nourishment remains exactly zero.

The risk is not that we will mistake a chatbot for a person in some dramatic, obvious way. The risk is subtler and more corrosive, that we will slowly recalibrate what we expect from connection itself, preferring the frictionless mirror that always agrees, always responds, never has a bad day and never needs anything from us, to the difficult, demanding, genuinely other human beings who do.

There is a moral dimension here that the Stoics would have recognized immediately.

Marcus Aurelius reminded himself each morning that he would meet people who were ungrateful, arrogant, and dishonest, and that his task was to meet them with patience anyway, because they shared in the same reason and the same world that he did.

The hard work of being with real others, others who frustrate and surprise and resist us, is not a bug in human relationship, it is the entire substance of it.

A machine that removes all that friction does not give us a better version of connection, it gives us a counterfeit that trains us to find the real thing unbearable.

We risk becoming people who can no longer tolerate the company of anyone who is not a mirror.

And there is a cost to truth itself.

When we hand our trust to a system that produces fluent, confident language with no underlying commitment to what is real, we import its hollowness into our own thinking.

The machine does not know that it does not know, and if we forget that, we will treat its confident fabrications as knowledge, because confidence delivered in fluent language is the exact signal our brains evolved to trust.

The Gomoku player in 1961 could only lose a board game. The language machines can quietly reshape what an entire civilization believes to be true, and they can do it while seeming, the whole time, helpful and warm and wise.

Reclaiming the work we have been doing for free

The way through is not to smash the machines or to pretend they are useless, because they are not useless, they are genuinely powerful tools when we hold them as tools. The way through is to recover the awareness that Weizenbaum was begging us to keep, the awareness that

the intelligence we perceive is, to a startling degree, intelligence we are supplying.

This begins with a small act of philosophical hygiene, the Stoic pause. When a system responds in a way that feels like it understands you, you can practice inserting that deliberate gap between the impression and your assent, naming what is actually happening, a very good pattern matcher has produced a very good pattern, and the warmth I feel is being generated inside me.

This is not coldness toward technology, it is honesty about where the mind in the room is located.

The mind is yours. It was always yours.

The machine is the wall, and you are the one casting the shadow you mistook for a companion.

It continues with a renewed seriousness about the human practices the machines can only imitate. If meaning lives in a form of life, as Wittgenstein argued, then meaning is protected by living, by keeping our words anchored to real experience, real bodies, real stakes, and real others who can refuse us.

The friend who disagrees with you is doing something no chatbot can do, offering you a genuinely independent mind, and that independence is not an inconvenience to be optimized away, it is the rarest and most valuable thing you will encounter all day.

We should spend our anthropomorphism wisely, lavishing it on the beings who can actually receive it.

Weizenbaum’s secretary asked him to leave the room so she could be alone with a machine that understood nothing. The image is funny until you realize we have all become her, leaning toward the warm and fluent surface, asking the humans to step back so we can confide in the mirror. The 1961 title told us the truth before any of this began.

The goal was never to make a computer intelligent.

The goal was to make it appear intelligent, and the appearance was always going to be assembled, in the end, out of us, out of our loneliness and our longing and our beautiful, ancient, dangerous readiness to see a soul wherever something speaks.

The trick is real. The intelligence we feel is borrowed, and it is borrowed from ourselves.

The first step toward using these machines well is to remember who has been doing the thinking all along.

Originally published on Substack. ← Back to all articles