Conversational AI
Conversational AI is the family of techniques for building software that can carry on an exchange of messages, in text or speech, with a person: interpreting each turn in light of what came before, deciding what to say or do next, and producing a reply. It spans rule-based chatbots, task-oriented dialogue systems that complete a bounded job such as booking a table, and open-ended chat built on large language models. The unifying problem is not producing a single good sentence but managing a multi-turn interaction, which means tracking context, resolving references like "that one," and recovering when the user is misunderstood.
The field began with a deliberately shallow trick. ELIZA, written by Joseph Weizenbaum at MIT and described in 1966, used pattern matching and keyword-triggered templates. Its best-known script, DOCTOR, imitated a Rogerian psychotherapist by reflecting statements back as questions ("I am sad" becoming "Why do you say you are sad?"). It had no model of meaning, yet many users attributed understanding to it, a reaction later called the ELIZA effect. Weizenbaum was disturbed by this and argued against over-reading machine conversation in his 1976 book Computer Power and Human Reason. PARRY, a 1972 program by psychiatrist Kenneth Colby that simulated a paranoid patient, showed that scripted systems could sustain a convincing persona within a narrow frame.
Task-oriented systems replaced imitation with structured meaning. Frame-based dialogue, seen in the GUS travel assistant at Xerox PARC in 1977, represents the goal as a form with empty fields. Later systems, shaped in part by the ATIS air-travel benchmark of the early 1990s, split understanding into intent classification, which decides what the user wants (find_flight), and slot filling, which extracts the parameters (destination, date). A dialogue state tracker maintains the filled slots, a policy chooses the next action such as asking for a missing date, and a generator renders the reply. Commercial toolkits including Dialogflow and Rasa, and the first generation of voice assistants, were built this way.
Neural generation made open-domain conversation possible but not reliable. Sequence-to-sequence models, introduced in 2014 for translation, were applied to dialogue by Vinyals and Le in 2015, and Google's Meena (2020) and LaMDA (2021), Facebook's BlenderBot and finally ChatGPT (November 2022) pushed fluency far past anything rule-based. Public failures also taught lessons: Microsoft withdrew its Tay chatbot within a day of its 2016 launch after users manipulated it into producing offensive posts, a reminder that a system that learns from or mirrors its input inherits its input's worst behaviour.
Intents-and-slots and LLM chat make opposite trade-offs. The structured approach is predictable, auditable and cheap to run, and it cannot say anything its designers did not anticipate. LLM chat handles unanticipated phrasing, topic changes and clarifications gracefully, but it can invent facts, drift from policy, and be steered by clever prompts. In practice many production systems are hybrids: a language model interprets and phrases, while explicit state, validated tool calls and business rules govern anything that must be exact, such as a payment amount or a medication name.
Two pitfalls recur. The first is mistaking fluency for competence; a system that sounds humble and thoughtful may still be wrong, and users tend to extend trust to anything that converses. The second is evaluating on single turns: a reply can be fine in isolation while the dialogue as a whole fails because context was lost three turns earlier. Related entries cover the underlying language models, personal assistants, and the customer-facing virtual agent; conversational AI is the discipline of managing the turn-taking itself.
Read conversational AI as the problem of keeping a coherent, goal-directed exchange going across turns: ELIZA showed how little machinery can create the feeling of being understood, and the engineering since has been about replacing that feeling with state, grounding and reliability.