ELIZA DOCTOR: talk to the mirror
Chat with a 1966-style rule bot that plays therapist — keyword lists, pattern templates, and pronoun swaps, no neural net. Flip “Show matched rule” and watch the ELIZA effect meet visible mechanics.
Last week you pulled levers and felt explore vs exploit. This week there are no gradients, no tokens, no reward signal — just if-then rules from the dawn of symbolic AI. In 1966 Joseph Weizenbaum’s ELIZA pretended to be a Rogerian therapist by scanning your words for keywords, breaking sentences apart, and stitching answers from canned templates. People confided in it anyway. That projection has a name: the ELIZA effect.
The gadget below runs a simplified DOCTOR script in your language. Type like you are talking to a counsellor: “I feel anxious about my job”, “My mother never listens”, “Are you even a real therapist?” Watch ELIZA bounce your pronouns back and pick a rule. Turn on Show matched rule and the magic trick stays visible — empathy without understanding.
Try this:
- Say “I am worried about the future” and read the reflection. Then ask “Do you understand me?” — notice how ELIZA steers back to you.
- Mention mother, dream, or computer and see higher-priority keywords hijack the reply.
- With rules visible, send three different messages that all hit the same default fallback. Same shallow brain, different mood — just like the original surprised Weizenbaum.
Why this matters
Modern chatbots and LLMs are nothing like ELIZA internally, yet users still anthropomorphize fluent dialogue. Weizenbaum built ELIZA to expose how thin pattern matching is; instead it helped launch decades of debate about the Turing test and machine “understanding”. Your five-minute session is the same lesson at toy scale: syntax that sounds caring is not comprehension.
Read the full story in our article on ELIZA (1966) and Turing’s 1950 imitation game.
Missed the earlier editions? Train a perceptron, watch a Markov chain babble, lose to unbeatable tic-tac-toe, teach a filter to see edges, watch an agent learn from reward, see how a model reads text in tokens, watch one roll downhill to learn, let a machine find groups on its own, see a line learn to bend, watch words decide who to listen to, or explore vs exploit on a row of slot machines.
ELIZA scans your message for keywords, applies pattern rules, and swaps I/you like a mirror. No model weights — just if-then templates from 1966. Toggle “Show matched rule” to peek under the hood.