ELIZA: The First Chatbot

In 1966 at MIT, Joseph Weizenbaum wrote ELIZA, a program that mimicked a Rogerian psychotherapist. You typed a sentence, and ELIZA rephrased it back as a question. Simple pattern matching, no understanding at all.

If you typed I am sad, ELIZA might reply, why do you say you are sad? If you mentioned your mother, it would ask about your family. The trick worked because Rogerian therapy is mostly about reflecting the patient back to themselves.

How ELIZA worked

1. 1Scan the user's input for keywords
2. 2Match the keyword to a pattern rule
3. 3Rephrase parts of the sentence using that rule
4. 4When stuck, ask a generic prompt like, tell me more

Weizenbaum grew alarmed at how readily people anthropomorphized his toy. In 1976 he wrote Computer Power and Human Reason, a book warning against putting machines in roles that require human judgment and compassion.

“I had not realized that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people.”

The big idea: humans are wired to find minds in language. That makes building helpful AI powerful and building honest AI hard.