Chatbot
A program you talk to in ordinary language — and the oldest demonstration that people credit machines with understanding on almost no evidence.
Reviewed July 16, 2026Stable
When not to use it
- Lookup problems. If users want one answer, a search box or an FAQ beats a conversation and doesn't hallucinate.
- Anything with a legal or financial commitment at the end. Fluent and wrong is a liability, and it's the default failure mode.
- Where users will type "agent" immediately. Some interactions should be a form or a person, and burying that behind a chat is a hostile design.
Reach for something else instead
- A good FAQ or search page solves most of what chatbots are deployed for, faster and with no latency.
- Structured forms are better anywhere the required information is known in advance.
- A person remains the right answer more often than anyone building one wants to hear.
Sources & further reading
- Weizenbaum (1966), ELIZA — A Computer Program for the Study of Natural Language Communication Between Man and Machine — the original, and the first documentation of the effect named after it.
- Weizenbaum (1976), Computer Power and Human Reason: From Judgment to Calculation — the author's book-length objection to what people concluded from ELIZA.
- Shanahan (2023), Talking About Large Language Models — the careful modern treatment of what our vocabulary for these systems smuggles in.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Deploying conversation onto a lookup problem. Users don't want to chat with your bank; they want a balance.
- Assuming fluency is understanding. That inference was wrong in 1966 with 200 lines of pattern matching, and fluency has improved much faster than the grounds for the inference.
- Losing the visible failure. Old bots said "I don't understand"; modern ones produce a confident wrong answer, and only your evals will ever notice.
At a glance
FieldApplied AI
FirstELIZA, Weizenbaum, 1966
Named effectthe ELIZA effect
Modern failurefluent and wrong, invisibly
DifficultyBeginner