A later ELIZA demonstration shows how the program turns parts of a person’s message into a reply. Image source. Licence.
Tell a computer about a family member and it can ask you to say more about your family. Joseph Weizenbaum's ELIZA produced replies like that by looking for words or sentence patterns and applying a script. His 1966 paper showed how the process worked.
A script could break a sentence into parts and put some of those parts back together in a different order. That let ELIZA turn something a person said into a question that appeared to follow the conversation. Its best-known script imitated a form of psychotherapy, a setting in which asking someone to elaborate can sound natural even when the reply contributes little new information.
The script did much of the work
Weizenbaum separated the program from the scripts that supplied its conversational behavior. Change the script and the same program could give a different sort of reply. The paper's examples make the effect inspectable. Readers can trace how the text a person typed became the text the machine returned.
Today's language models generate replies differently. ELIZA's lesson survives that change in machinery. A conversation can feel attentive because of what the interaction invites the person to supply.