chatbot

In depth

Two families

The word predates language models by several decades, and it now covers two very different families. The first runs through a decision tree written in advance: recognised words, prepared answers, perfectly predictable behaviour as long as one stays within the script, and a dead end as soon as one leaves it. The second entrusts the answer to a model that reads the exchange and writes: coverage becomes very broad, at the price of a share of unpredictability. From the outside the two look alike, for a thread of messages says nothing about what lies behind it.

Three words to keep apart

It is useful to separate three words that are often used one for the other. Chatbot names the form of the exchange, assistant names the role held alongside you, agent names a way of working towards a goal. One and the same system can be all three at once, but nothing obliges it to be: an agent can work without any conversation existing, and a chatbot can be no more than a form in disguise. Adding tools or a documentary base enriches the harness without shifting this question: as long as the system waits for your message before acting, the conversation is led by you.

The apparent memory

A chatbot remembers nothing by itself: the apparent continuity comes from the harness, which puts the history back into the context at every turn, and it stops where the harness has decided that it stops. The trap in evaluation is to judge the fluency of the exchange rather than the work accomplished: a pleasant system that has not settled the request has settled nothing. The trap in design answers it as a mirror image, for a conversation is a costly and imprecise way of gathering three pieces of information that a form would obtain without ambiguity. Finally, the conversational interface hides its own limits, since it always answers something: guardrails and the possibility of admitting ignorance count here more than elsewhere.