What was the first chatbot, before LLMs existed?

The first well-known chatbot was ELIZA, built in the mid-1960s by Joseph Weizenbaum at MIT — decades before anything resembling an LLM existed. It imitated a psychotherapist by matching keywords and echoing them back as questions, and it was convincing enough that people genuinely felt understood.

How did ELIZA actually work?

ELIZA didn't understand language at all. It ran a simple script called DOCTOR that pretended to be a psychotherapist: it spotted keywords in what you typed and reflected them back as questions. Type "I'm angry at my mother" and it might reply "Why do you feel angry at your mother?" No reasoning, no memory of the conversation, just pattern-matching and templates — yet plenty of people using it got genuinely convinced they were talking to something that understood them.

Weizenbaum himself was unsettled by how readily people confided in it, a reaction now known as the ELIZA effect: our habit of reading real understanding into a machine that has none.

What other chatbots came before LLMs?

ELIZA wasn't alone. A few landmarks from the rule-based era:

• PARRY (1970s) — built by psychiatrist Kenneth Colby, it simulated a person with paranoid schizophrenia, rule-based in the same spirit as ELIZA.
• Decision-tree bots (2000s-2010s) — the rigid support bots on company websites and phone menus that understood only a handful of preset phrases.

Stray off the script with one of those website bots and you'd get looped straight back to "I'm sorry, I didn't understand that."

How are these different from ChatGPT?

All of the early bots share one thing: they were hand-written rules and pattern-matching, not learning. When LLM-based chatbots like ChatGPT showed up in the 2020s, they looked similar on the surface — a text box, a back-and-forth conversation — but underneath they're a completely different technology, trained on huge amounts of text to actually generate novel language rather than follow a script. ELIZA could only recombine your own words; an LLM can produce a sentence no one wrote in advance, which is why it can hold a real conversation instead of bouncing you around a menu.