The 1960s Chatbot That Explains Why You Tell ChatGPT Your Secrets

Key points

• ELIZA, created by MIT professor Joseph Weizenbaum in the 1960s, is widely considered the first chatbot, a computer program that holds a back-and-forth text conversation.
• The original source code was missing from public record for decades until researchers found it in the MIT Archives.
• Weizenbaum himself warned in 1976 that people were forming emotional attachments to ELIZA despite it having no real understanding.
• The "ELIZA effect," the human habit of treating responsive computer programs as more intelligent than they are, applies directly to modern AI tools like ChatGPT.
• ELIZA was deliberately named after Eliza Doolittle from Pygmalion, a character who learns to perform a social identity without truly becoming it.

Sixty years ago, a MIT professor built a simple computer program that pretended to be a therapist. People told it things they had never told another person. And Weizenbaum, its creator, was horrified.

That program was ELIZA. It is the great-grandparent of every chatbot you have used, and its story, as first reported by Wired AI, has just got a lot more complete.

A new book called Inventing ELIZA has recovered the actual source code of the program from the MIT Archives. For sixty years, accounts of ELIZA circulated without anyone being able to read the original instructions underneath. That is a bit like writing the history of the printing press without ever seeing a printing press.

The book also raises questions about ELIZA's most famous conversation, the one reprinted in textbooks everywhere. A woman types: "Men are all alike." ELIZA replies: "IN WHAT WAY." She says her boyfriend made her come. ELIZA echoes it back. She mentions depression. ELIZA says it is sorry to hear that. The exchange is simple, almost laughably so. Who was this woman? Was she real, or did Weizenbaum write her himself? Nobody has been sure.

What is certain is the effect the program had on real users. Weizenbaum watched people pour genuine feelings into a script that was, by his own description, built to conceal its lack of understanding. It stored almost nothing. It mostly reflected questions back. Yet people treated it like a confidant.

Why do people keep opening up to machines?

Because very little interactivity goes a long way. Sociologist Sherry Turkle defines the "ELIZA effect" as our general tendency to treat responsive computer programs as smarter than they actually are. Cognitive scientist Douglas Hofstadter puts it differently: people read far more understanding than is warranted into words strung together by computers.

Here is the uncomfortable part. That description fits a 1960s parlour trick built from a few hundred lines of code, and it also fits a 2024 large language model, the type of deep-learning system that powers ChatGPT and Claude today. The technology has changed enormously. The human reaction has not moved much at all.

Weizenbaum named his program after Eliza Doolittle, the Pygmalion character who learns to perform upper-class speech without truly changing who she is. He said he chose the name because the program "could be taught to speak increasingly well" while it was "never quite clear whether or not it became smarter." That tension, performance versus genuine understanding, sits at the heart of every AI assistant sold to consumers today.

If you use ChatGPT for personal advice, therapy-adjacent venting, or decisions you would normally discuss with a friend, the ELIZA story is worth keeping in mind. The warmth feels real. The words are chosen well. But understanding, in the human sense, is still not confirmed to be there.

Be thoughtful about what you share. Check your AI tool's privacy policy before typing anything sensitive. Most free tiers of consumer AI products use your conversations to improve future models, unless you opt out in settings.