The Original Chatbot Had Many Faces, and We Never Knew It

Key points

• ELIZA, created by Joseph Weizenbaum and first published in January 1966, is widely considered the world's first chatbot.
• Researchers found ELIZA's original source code sitting untouched in MIT's archives, years after reconstructions based only on published descriptions became standard.
• The code reveals ELIZA was not a single program but a platform that could run multiple distinct personas, or "scripts", covering topics from poetry to elevators to the French language.
• A new book, Inventing ELIZA: How the First Chatbot Shaped the Future of AI, published by MIT Press in 2026, details the discovery and lets readers try a working emulation of the original.
• The gap between what Weizenbaum published about ELIZA and what the code actually does mirrors a tension that still runs through AI development today.

Picture a computer program from 1966 that could hold a conversation. Not a clunky yes/no exchange. A genuine back-and-forth, where the machine asked follow-up questions and seemed to actually listen.

That program was ELIZA, and it startled everyone who met it.

Its creator, MIT computer scientist Joseph Weizenbaum, built it as an experiment. Users would type their worries, and ELIZA would respond in the gentle, questioning style of a therapist. People found it so convincing they shared real feelings with it. Some believed it understood them. Weizenbaum was alarmed by his own creation.

The phenomenon got a name: the ELIZA effect, the tendency to assume a computer has genuine feelings or understanding when it is simply following rules.

Here is the twist. For sixty years, nobody was working from the real code. Researchers rebuilt ELIZA from Weizenbaum's published 1966 paper in the journal Communications of the ACM (the Association for Computing Machinery, a major professional body for computer scientists). That paper, all ten pages of it, left out key details.

The actual source code sat in MIT's archives until very recently.

When researchers finally dug it out, they found something surprising. ELIZA was not one program. It was a platform, a kind of engine that could run different "scripts", each one giving the system a completely different persona.

The famous therapist character? That was just one script, called Doctor. Weizenbaum chose psychiatry as the persona because a therapist can plausibly know almost nothing about your real life and still hold a conversation. It was a clever workaround for the program's limits, not a statement about mental health.

Other scripts let ELIZA discuss mathematics, poetry, colour, paradoxes, relativity, France and, brilliantly, elevators.

As IEEE Spectrum AI first reported, the new book Inventing ELIZA frames this as a genuine turning point in understanding early computing history. The code shows ELIZA had contextual memory, the ability to track earlier parts of a conversation, and script-editing features nobody knew about.

Does any of this matter for modern AI?

Yes, more than you might expect. The design choice to separate the ELIZA engine from its scripts, keeping the rules for conversation in a swappable file rather than baked into the program itself, directly anticipated how modern AI chatbots work. Today's systems, from customer-service bots to voice assistants, run different "personas" on top of shared underlying models. ELIZA did that first.

The gap between the published description and the real code is also telling. Weizenbaum explained some of what ELIZA did but not all of it. That distance between the theory and the actual implementation, between what developers say a system does and what the code really does, is a live issue in AI today.

For ordinary users, the lesson is simpler. The program that made millions of people feel heard was always more complicated, and more limited, than it looked. That is still true of the chatbots on your phone.

ELIZA never understood a word anyone said. It just knew how to make you feel like it did.