Last week, while preparing a presentation for my PhD studies on AI-human connections and human dependence on technology, I fell down a research rabbit hole that caught me completely off guard. Everyone who has had this conversation in the past few years has started it at the same place: ChatGPT, released in late 2022. The moment the public discovered that a machine could hold a conversation, could listen patiently, could respond in ways that felt — somehow, uncomfortably — like being heard.

But the more I dug, the further back the story went. Past Alexa. Past Siri. Past the early chatbots of the 1990s. All the way to 1966, to a dimly lit computer lab at MIT, and a program called ELIZA — where the exact same thing had already happened, nearly sixty years earlier. And where its creator, rather than celebrating, was horrified.

A program that was never supposed to feel real

Joseph Weizenbaum built ELIZA between 1964 and 1967. It was not designed to be intelligent. It was designed to demonstrate the opposite: that the appearance of a meaningful conversation between a human and a machine was, in Weizenbaum’s word, an illusion. A parlor trick. A mirror that reflected your own words back at you, rearranged just enough to feel like a response.

Its most famous version ran a script called DOCTOR, modeled loosely on a Rogerian psychotherapist — the kind that responds to what you say by gently asking you to say more of it. You would type something like “I feel like no one understands me,” and ELIZA would respond: “Why do you think no one understands you?” No comprehension was taking place. No understanding. ELIZA was doing pattern-matching and grammatical substitution. It had no interior life. It had no idea what you were talking about.

User: Men are all alike.
ELIZA: In what way?
User: They’re always bugging us about something or other.
ELIZA: Can you think of a specific example?
User: Well, my boyfriend made me come here.
ELIZA: Your boyfriend made you come here?
User: He says I’m depressed much of the time.
ELIZA: I am sorry to hear you are depressed.

— Verbatim exchange from Weizenbaum, J. (1966). “ELIZA — a computer program for the study of natural language communication between man and machine.” Communications of the ACM, 9(1), 36–45. 

Weizenbaum published his paper in January 1966 in Communications of the ACM. He expected people to see through it immediately. He expected the illusion to shatter on contact.

It didn’t.

The secretary incident

The first sign that something unexpected was happening came from inside Weizenbaum’s own office. His secretary — someone who knew full well that ELIZA was a program her boss had written himself — sat down at the terminal one afternoon and began to use it. After a short time, she asked Weizenbaum to leave the room.

She wanted to be alone with it.

Weizenbaum was shaken. He later wrote: “I had not realized… that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people.”

He had not expected normal, educated, clear-minded people to project emotional depth onto a program they knew was hollow. He had not expected them to feel listened to, to feel understood, to feel that something genuine was happening on the other side of the screen. But they did. They opened up. They shared things they hadn’t shared with other people. And the more they shared, the more the illusion deepened.

This phenomenon — the human tendency to attribute understanding and feeling to conversational software — eventually acquired a name: the ELIZA effect. We are apparently wired to read intelligence into language. When something responds to us in sentences, we respond to it as if it is someone.

A creation he wished he’d never made

The reaction from the broader research community made things worse, not better. Rather than sharing Weizenbaum’s concern, many psychologists and AI researchers were excited. Some openly proposed that ELIZA — or programs like it — could be used as a form of automated psychotherapy. They were serious. Weizenbaum was appalled.

In 1976, a decade after ELIZA’s release, he published Computer Power and Human Reason: From Judgment to Calculation. It is a difficult, urgent, at times almost anguished book. Weizenbaum was not arguing that computers were unintelligent. He was arguing that there were things only humans should be trusted to do — things that required empathy, care, wisdom, moral judgment — and that delegating those things to machines, no matter how convincingly they mimicked understanding, was not a technical problem but a moral failure.

He was specifically haunted by the idea that people might come to prefer the company of a program to the messiness of actual human connection. That the machine’s availability, its patience, its perfect attention, could become a substitute for the real thing — and that this substitution might feel like an improvement.

What Weizenbaum was actually worried about

Almost sixty years on, preparing a presentation on exactly this question — AI-human connection, and the psychological dependencies these systems generate — I keep coming back to a specific detail. Not the technology itself, which has changed almost beyond recognition. But the experience that Weizenbaum witnessed: of people feeling understood, not by someone, but by something, and of that feeling being genuine enough to make them want the room to themselves.

ChatGPT is not ELIZA. The systems of 2025 produce language at a level of fluency that ELIZA could never approach. They can sustain coherent conversations for hours. They reference earlier parts of a conversation. They shift tone. They express — convincingly — something like curiosity, care, concern. The gap between the two technologies is enormous.

But the gap in what they produce in the people using them may be smaller than it first appears. The same reaching for connection is there. The same sense of being listened to. The same willingness to share things you might not share elsewhere. The ELIZA effect hasn’t gone away — it has simply found far more capable surfaces to bounce off.

Weizenbaum was not opposed to computers. He was not opposed to progress. He was worried about what happens when a technology that simulates intimacy encounters a species wired to feel intimacy in response to it. And he was worried about what we lose — what capacity for tolerating the difficulty of real human connection — in the process.

He died in 2008, twenty years before ChatGPT. I suspect he would have had thoughts about the moment we are living through. I suspect they would not have been simple, and they would not have been comfortable.

That, in itself, seems worth thinking about.