From ELIZA to ChatGPT: How humans continue to bond with machines that seem to understand them

Long before ChatGPT, an employee at the Massachusetts Institute of Technology (MIT) sat down at a terminal and began chatting with a computer program. She had watched her boss develop the code for months and knew exactly what it was. Yet, after exchanging a few sentences with the machine, she turned to it and asked, “Would you mind leaving the room, please?”

The boss was Joseph Weizenbaum, and the program was called ELIZA. First presented in a 1966 scientific article, it is considered the world’s first chatbot. Sixty years later, the secretary’s story has become a legend in computer science and remains the best illustration of a phenomenon that artificial intelligence companies have yet to solve: how easily people bond with machines that seem to understand them.

The name came from Eliza Doolittle, the florist in Bernard Shaw’s play Pygmalion, which inspired the musical My Fair Lady. In the story, the protagonist learns to speak like a lady without changing her essence.

The program worked similarly: it identified keywords in the user’s sentences and returned them in the form of a question. Using a script called DOCTOR, it mimicked a psychotherapist who, instead of providing answers, encouraged the patient to keep talking. If someone wrote that they were sad, the machine would ask why they felt that way.

The trick worked. “Some participants were very difficult to convince that ELIZA was not human,” Weizenbaum wrote in his 1966 article. People opened up, shared intimate details, and treated the program as someone who could listen to them. The tendency to attribute feelings and intelligence to a machine became known as the “ELIZA effect.”

The inventor who became a critic

The reaction deeply troubled Weizenbaum.

Born in Berlin in 1923 to a Jewish family that fled Nazi Germany when he was a teenager, the scientist created ELIZA to explore communication between humans and machines, not to replace therapists. Many colleagues, however, saw the program as exactly that.

Psychiatrist Kenneth Colby of Stanford University adapted the idea into a program called PARRY, which simulated a patient with paranoid schizophrenia. In 1975, astronomer Carl Sagan even envisioned networks of booths, similar to public telephones, where people would pay a few dollars to talk to a computerized therapist.

Weizenbaum went in the opposite direction. In 1976, he published the book Computer Power and Human Reason, in which he recalled the secretary’s episode and issued a warning. “What I had not realized was that extremely brief exposures to a relatively simple computer program could induce strongly delusional thinking in quite normal people,” he wrote.

Until his death in 2008, he dedicated himself to questioning the idea that computers should make decisions that depend on human judgment.

The warning seems timely. The 1966 program occupied only a few lines of code; current models are trained on a large portion of what has been written on the internet and converse with more than 1 billion people per week on ChatGPT alone.

The technology has changed beyond recognition. The human reaction, not so much. Even today, some people use chatbots as therapists, confidants, and even romantic partners, and researchers are discussing the effects of this emotional dependence, especially among young people who spend more and more time online.

The difference is one of scale. Weizenbaum was startled by a secretary who asked for privacy to talk to a program she knew was just a program. Sixty years later, the same scene is repeated daily with hundreds of millions of people.