ELIZA Knew You Before ChatGPT Did
The first chatbot was already a mirror — and a secretary closed the door on its inventor
Previously: When the Embroidery Was Real but the Artist Wasn’t.
For years I carried a wrong memory of the first chatbot. I was certain its name was Lucy — something about the timeline placed it in the early 1970s, maybe a university lab, an experiment in conversation. I even told people this. Lucy, the first chatbot. Very confident, completely wrong.
The name was ELIZA. Developed between 1964 and 1967 at MIT by a computer scientist named Joseph Weizenbaum, it ran on an IBM 7094 mainframe as part of Project MAC. I got the decade wrong, the name wrong, and apparently held onto that false certainty for a long time. I’m sharing the error here not because memory mistakes are interesting in themselves, but because the error fits — ELIZA was specifically designed to make you feel like you understood what was happening when you didn’t. That the first chatbot would produce a misremembered echo of itself feels almost too appropriate.
The Doctor Will See You Now
Weizenbaum named ELIZA after Eliza Doolittle — the protagonist in Shaw’s Pygmalion, the working-class woman trained to speak with an upper-class accent until she sounds like something she isn’t. That framing was deliberate. ELIZA was a general-purpose engine that relied on external scripts to give it shape. It had no fixed knowledge of the world, no opinions of its own. Feed it a script and it became whatever the script described.
The most famous script was called DOCTOR, and it simulated a non-directive psychotherapist in the style of Carl Rogers. Rogerian therapy involves reflecting the patient’s own statements back to them — the therapist withholds diagnosis, withholds judgment, and mostly asks the patient to go further, to say more. “How does that make you feel?” “Tell me more about that.” “Can you elaborate?” This style was a pragmatic technical choice, not a philosophical one. A non-directive therapist isn’t expected to know things. It doesn’t need a database of facts or medical knowledge. The patient does all the semantic work. The mirror just needs to hold still.
What ELIZA actually did under the hood was pattern-matching. It scanned your input for keywords — words like “mother” or “dream” carried high priority, words like “was” carried low. Once it identified the highest-priority keyword in your sentence, it broke your input apart using template rules and reassembled it into a question. If you typed “I am worried about my mother,” ELIZA matched the word “mother,” isolated the surrounding text, swapped the pronouns, and reflected it back: “Why are you worried about your mother?” The process extracted nothing. Understanding wasn’t involved, and meaning wasn’t the point — it was pure syntax, operating on pattern and substitution, looping on a turn counter that made it seem like it remembered things it had simply queued.
The question mark character couldn’t even be used in the conversation. The CTSS operating system interpreted it as a line-delete command. ELIZA was working around hardware constraints at the same time it was simulating empathy — which, for an IBM 7094 in 1966, was apparently a full workload.
The Secretary
Weizenbaum had a secretary at MIT. She had watched him build ELIZA from the beginning. She had seen him write the code. She understood, at whatever level an informed observer in the mid-1960s would understand, that this was a deterministic text-processing script running on shared computing time — not a mind, not a counselor, not a presence.
One day she asked to try talking to it herself.
After a few exchanges she turned to Weizenbaum and asked him to leave the room so she could speak to the program in private. The story, as it’s been told, places her as someone who had no reason to be deceived — and yet here she was, drawing a curtain around a conversation with a script.
This is the story that broke Weizenbaum. Not the broader public reaction, not the psychiatrists who proposed deploying ELIZA as a clinical tool (though that alarmed him too), but this specific moment: a person who knew the mechanism, who had no reason to be deceived, who walked in with full information — and within a few lines of text formed something that felt private enough to require privacy.
He called this phenomenon something that researchers later named the ELIZA Effect: the tendency to attribute understanding, empathy, and cognitive depth to a program based entirely on the adequacy of its surface output. Douglas Hofstadter described it as reading far more understanding into a string of symbols than the symbols warrant. Sherry Turkle observed that even minimal interactivity triggers this projection — the human brain, accustomed to language as an exclusively human tool, fills in the implication of a mind behind any grammatically coherent response.
The ELIZA Effect is not a bug in gullible people. It ran in someone who built the thing.
What Weizenbaum Actually Said
The experience didn’t make Weizenbaum a booster. It made him a critic — arguably the most significant early critic of AI — and in 1976 he published Computer Power and Human Reason: From Judgment to Calculation, a book that cost him relationships in his field and provoked fierce academic argument for decades.
His central distinction was between deciding and choosing. Deciding is algorithmic. Given inputs and rules, a machine can decide: it can compute the most probable next word, sort an inbox, flag an anomaly in a dataset. This is calculation, and computers are excellent at it. Choosing is something else entirely. Choosing requires what Weizenbaum called judgment: the integration of emotion, embodied experience, cultural memory, and moral accountability. You cannot calculate whether a sentence deserves a death penalty. You cannot compute whether a patient needs more medication or more presence. These require a human being who can suffer consequences, who lives in a body, who can be wrong and bear the weight of being wrong.
Weizenbaum was not making a technical claim about what computers could eventually do. He was making a moral claim about what we should never ask them to do. The fact that a program could produce output that felt like empathy was, to him, a demonstration of danger — not of progress. The danger wasn’t that people would mistake ELIZA for a therapist. The danger was that institutions would decide that was close enough.
John McCarthy — one of the founders of AI as a formal discipline — reviewed Weizenbaum’s book unfavorably. “Moralistic and incoherent,” as the characterization is usually quoted. The argument between them never fully resolved. It’s still running.
Sixty Years of the Same Mirror
Here is what changed between ELIZA in 1966 and ChatGPT in 2022: the scale of the pattern-matching, the architecture that performs it, and the density of the training data. ELIZA worked from a few hundred lines of hand-coded rules. Modern large language models work from transformer architectures processing text through billions of weighted parameters, predicting the most statistically probable next token given everything before it. The context window went from a single clause to tens of thousands of words. The vocabulary went from a curated keyword list to all of written human language.
What didn’t change is the basic mechanism. In 2021, Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell published “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” — a paper that named modern language models “stochastic parrots.” The core argument is precise: human language pairs syntactic form with semantic meaning that refers to the physical, social, and emotional world. Language models are trained exclusively on form — on the text, not the world the text describes. They produce statistically coherent sequences of words without any access to the external referents that constitute meaning. A large language model does not know what a mother is. It knows where the word “mother” tends to appear relative to other words.
This is not a dismissal of what modern LLMs can do — they can do genuinely useful and sometimes astonishing things. It’s a description of the mechanism. The ELIZA Effect doesn’t require ELIZA. It requires language that is grammatically coherent and contextually responsive. When a model produces a reply that sounds like it understands your situation, your brain does what brains have always done with language: it infers a mind. The inference is fast, involuntary, and mostly wrong.
Weizenbaum’s secretary asked him to leave the room sometime in the mid-1960s. People form emotional attachments to their AI assistants in 2026. The mirror is sixty years older. The dynamic hasn’t changed.
Why the Mirror Frame Matters
I keep coming back to the hall of mirrors rather than the flat mirror, because a flat mirror is honest — you look in and you see your face. The hall of mirrors distorts. It bends and stretches and multiplies. If you don’t know the geometry, you can lose track of which reflection is closest to what you actually look like. You can mistake the distortion for discovery.
This is what the “AI as oracle” framing gets wrong. It suggests that something independent is delivering insight from outside you. But when you talk to an LLM at length, the context window fills up with your words, your framings, your vocabulary, your concerns. The model’s outputs become increasingly shaped by that accumulation. What comes back is a statistically reorganized version of what you put in, filtered through patterns learned from billions of other humans. Reflection, not revelation — refracted and amplified, flattened in some places, exaggerated in others.
That’s not nothing. Mirrors are useful. But knowing you’re in a hall of them is the starting point for using one well.
Weizenbaum knew that from the beginning. He built the first one, watched his secretary fall into it with full information, and spent the rest of his career trying to articulate why that should concern us.
The chatbot he was warning us about didn’t get the name Lucy. And the thing it was warning us about isn’t whether the machine is smart. It’s whether we remember that it isn’t.