Chatbot

A chatbot handles interaction through conversation rather than through controls. The idea is old: ELIZA, written by Joseph Weizenbaum at MIT in 1966, simulated a therapist with simple pattern matching, and people attributed understanding to it that it did not have. That mismatch between apparent and actual capability is still the central design problem.

Why so many disappoint

A conversational interface offers an unbounded input to a bounded system. A text box implies you can ask anything, and the moment it fails on a reasonable request, trust drops sharply. Buttons never make that promise, which is why they fail more gracefully.

Large language models changed the capability side of this substantially, and did not remove the design problem. A system that answers fluently and is sometimes wrong is harder for a user to calibrate against than one that is obviously limited.

What helps

Show scope rather than implying omniscience. Offer suggested actions so the user learns what the system can do without guessing. Make failure states useful, ideally with a route to a human. And never claim more certainty than the system has.

In practice

The right test is not whether the conversation feels natural, but whether the user finished the task faster than the button would have. Often they did not.