A chatbot is a software program that conducts conversation with a person through text or speech, interpreting what it receives and producing a response in natural language.
Two lineages sit under the word. Rule-based bots — from 📝ELIZA in 1966 through the scripted support agents of the 2010s — match input against patterns or walk a decision tree and return authored responses. They are cheap, predictable, and brittle exactly at the edges of what someone thought to script. Model-based bots generate responses from statistical patterns learned over large bodies of text, which makes them fluent on topics nobody scripted and unreliable in ways nobody scripted either.
The commercial wave of the late 2010s badly over-promised. Messaging-platform bots were positioned as an interface that would replace apps, and most shipped as decision trees wearing a chat skin; the gap between the pitch and the product soured the category for years. Large language models changed the substrate rather than the framing — the interface is still a conversation, but the operative constraint moved from "what did someone script?" to "what will the model assert?"
The persistent design problem is calibration. Fluency signals a competence the system may not have, and the user supplies the understanding the program lacks — the same gap Weizenbaum identified in 1966, now operating at far greater fluency and scale.