How ELIZA works

ELIZA read a sentence, found its most important keyword, matched a pattern the script gave for that keyword, and filled in a reply template with the person's own words. No learning and no prediction: every step was a rule a person wrote.¹

One rule from the 1966 script

(REMEMBER 5
((0 YOU REMEMBER 0) (DO YOU OFTEN THINK OF 4)
(DOES THINKING OF 4 BRING ANYTHING ELSE TO MIND)
(WHAT ELSE DO YOU REMEMBER)

In the script, 0 means "any number of words," and a number in a reply means that part of the match. So in (0 YOU REMEMBER 0), part 4 is whatever followed REMEMBER, and DO YOU OFTEN THINK OF 4 hands it back.¹

A tiny ELIZA

The same idea in modern code, written for this book (Weizenbaum's own program was in MAD-SLIP). It is for reading; you don't need to run it.

// Each rule: a keyword's rank, a pattern, and replies that reuse what was captured.
const rules = [
  { rank: 5, pattern: /\byou remember (.*)/i, replies: ['Do you often think of $1?', 'What else do you remember?'] },
  { rank: 2, pattern: /\bmy (.*)/i, replies: ['Your $1?', 'Why do you say your $1?'] },
  { rank: 0, pattern: /.*/, replies: ['Please go on.', 'What does that suggest to you?'] },
];
// Turn "I" into "you" and "my" into "your" before the words are echoed back.
const swap = { i: 'you', me: 'you', my: 'your', am: 'are', you: 'I', your: 'my' };
const reflect = (s) => s.split(' ').map((w) => swap[w.toLowerCase()] ?? w).join(' ');

function eliza(sentence) {
  const rule = rules.filter((r) => r.pattern.test(sentence)).sort((a, b) => b.rank - a.rank)[0];
  const reply = rule.replies[Math.floor(Math.random() * rule.replies.length)];
  const captured = sentence.match(rule.pattern)[1] ?? '';
  return reply.replace('$1', reflect(captured).replace(/[.!?]$/, ''));
}

eliza('Well, my boyfriend made me come here.');
// "Your boyfriend made you come here?" or "Why do you say your boyfriend made you come here?"

On the web, you can talk to a version of ELIZA in the Arcade: stevesparks.com/arcade/eliza.