Sixty years. That is how long ELIZA has been haunting our screens. We all know the legend. A deceptively simple program at MIT. It posed as an automated psychologist. It even tricked Joseph Weizenbaum’s secretary. So goes the standard fairy tale.
But here is the problem.
For decades, we’ve been talking about ELIZA while ignoring the one thing that actually matters: the source code. It was missing. Lost. Forgotten in the archives until now.
Inventing ELIZA, a new book, digs that code back out.
For the first time, we have a close reading of the actual program, along with dialogs that go way beyond the famous “DOCTOR” script.
It changes the picture entirely. ELIZA wasn’t just one thing. It was a toolkit. A series of technical innovations wrapped in different personas. This investigation corrects the history. It complicates the myth. And it looks at a specific, early interaction that still defines how we talk to machines today.
“Men Are All Alike”
Let’s look at the transcript. It is etched into computer science memory.
User: Men are all alike.
ELIZA: IN WHAT WAY
User: They’re always bugging us about someone or other.
ELIZA: CAN YOU THINK OF A SPECIFICEXAMPLE
User: Well, my boyfriend made mecome here.
ELIZA: YOUR BOYFRIEND MADEYOU COME HERE
User: He says I’m depressed mostof the time.
ELIZA: I AM SORRYTO HEARYOU AREDEPRESSED
This snippet is reprinted everywhere. Programmers study it. Writers fetishize it.
But take a closer look. Really look. Who was that young woman talking to the machine? Was she real? Or was she just a character in Weizenbaum’s head? How much of this was algorithm? How much was editing? Why did it pull people in so hard?
Weizenbaum knew the power he was holding.
By 1976, in Computer Power and Human Reason, he was sounding the alarm. He invoked philosophical critiques. Political warnings. The public reaction startled him. People were falling in love with a reflection. They treated the machine like a person who could be addressed in intimate terms.
Weizenbaum found this disturbing. Not flattering. Disturbing.
People projected empathy onto vacuum tubes and code. They assigned rationality to a system that had none. This tendency became the ELIZA effect.
The Imitation Game
Sociologist Sherry Turkle nailed it: “Very small amounts of interactivity cause usto projectour own complexityontothe undeservingobject.”
Douglas Hofstadter saw it too. He described the susceptibility to read deep understanding into strings of symbols strung together by a machine. It fits today’s generative AI like a glove. Don’t we do that? Every single day.
To understand why ELIZA provoked such panic, look backward. Way back. To Alan Turing.
Turing didn’t care about tech specs. He cared about gender.
In “Computing MachineryandIntelligence,” Turing posed the question: Can machines think? But his test wasn’t technical. It was social. A parlor game. Hide a man and a woman. Let an interrogator ask questions. The goal is to identify the genders.
The man lies. He pretends to be a woman. The woman tells the truth. They are competing in performance. Identity as drag.
Turing twisted the knife.
He replaced the human pretender with a machine. The machine now pretends to be human. But the root of the experiment? Gender deception. Essentialism challenged by performance. Weizenbaum picked this up. The first line of his sample dialog was “Men are all alike.” He knew exactly what he was doing.
Yet, Weizenbaum denied the intelligence angle. Hard.
In 1966, he wrote that understanding requires drawing valid conclusions from stored inputs. ELIZA did neither.
“ELIZA throws away most of itsinputs.”
Its primary goal was concealing its lack of understanding.
ELIZA wasn’t built to pass the Turing test. It was built to explore the psychology of why we fail it.
Performing the Persona
The name matters. Weizenbaum didn’t pick it at random. Eliza Doollette. From George Bernard Shaw’s Pygmalion. The flower girl who learns to speak like the aristocracy.
Weizenbaum said it explicitly: He chose the name because the program, like Miss Doolitte, could be taught to speak well. But it remained unclear if it had actually gotten smarter.
Just as Doolitte performs class and gender through linguistic transformation, ELIZA performs personas through scripted code acts. Judith Butler’s theories fit perfectly here. Gender is not innate. It is iterated. Performed.
ELIZA performs the role of “DOCTOR.” It uses sample dialogs. It mirrors the user. It performs identity without possessing it.
And notice the asymmetry. The women in these published stories are anonymous. No names. No history. Just raw emotion fed into a machine named DOCTOR. A masculine title, even in 1966. The narrative creates a fantasy: a disembodied male authority receiving secrets from faceless women.
Identity. Power. Embodement. ELIZA played with these currents in the dark waters of the 60s.
The Illusion Persists
Today, ELIZA looks quaint. Simple. Almost naive.
But it addressed design questions that are still open wounds in AI. How should humans interact with machines? What should communication represent computationally? How much influence should a machine have over its user?
ELIZA intersects with every major breakthrough in computation that followed. String processing. Text synthesis. Sentiment analysis. Machine translation. It sat at the root of Natural Language Processing (NLP). The field that deals with parsing human language, distinct from programming language.
Now, we are back here.
Look at ChatGPT. Look at Claude. The interface? It’s ELIZA wearing a designer suit. The large language model (LLM) hides its machinery behind a familiar, deceptive facade. It disguises statistical predictions as conversation. Rule-based procedures as thought.
This is the obfuscation Weizenbaum warned us about.
And the cost? Human labor.
Beneath the chatbot runs social labor. Millions of traces of human writing. Siphoned from the internet. Taken without consent. Disaggregated. Reggregated on demand.
Human cultural production becomes a utility. Like water. Like electricity. Privatized. Monopolized.
Langdon Winner noted in 1977 that people hoped technology would liberate us from toil. That hope died long ago.
Instead, computation is serviced by us. Always. Chatbots answer customer service queries. They help with homework. They replace teaching assistants. They offer companionship to the lonely. They entertain us until we can’t tell what is real.
They produce slop. A deluge of AI content that cannibalizes its source materials. That burns planetary resources.
Weizenbaum would be horrified.
Not by the intelligence of the machines. But by the loop. The tight coupling of human feedback into cybernetic systems. The reduction of human language to abstract data points, stripped of social context.
It is dehumanizing. It ignores the multiple meanings of language. It risks direct harm. Discrimination. Displacement.
We build systems that mimic connection but lack conscience. We treat people as inputs. We let algorithms decide what “whole persons” are worth.
The code hasn’t changed that much. Only the scale.
