A rule-based chatbot made inspectable. It shows how pattern matching, templates, and conversational polish can feel meaningful without any language-model understanding — making the gap between visible mechanism and felt effect the thing to study.
Live preview · launch for the interactive version
Every ELIZA response traces back to a pattern: a keyword, a template, and a fill-in. The rule that fired is right there on screen — before the response, not after.
Even with the rules visible, the conversation can feel caring or knowing. That gap — between what the mechanism does and what the output implies — is the central thing to name.
ELIZA’s rules show what a modern language model hides. When you compare them side by side, the opacity of the LLM becomes visible by contrast, not by description.
Don’t just chat with it — run it like an experiment. Predict what rule will fire, then check.
Write down what you expect ELIZA to do with a specific input. Which keyword will it match? Which template will fire?
Add, remove, or rephrase the triggering keyword. Does the matched rule change? Does the response change?
Run the same input through ELIZA and a pre-generated LLM example. What does each response reveal about what the system understood?
Name the specific gap: what does the visible rule not do that the LLM response appears to do? The LLM’s gap is just harder to see.