Autoregressive
The standard way language models write: one token at a time, left to right, each one chosen after the one before it.
Almost every chatbot you have used is autoregressive. It writes the way you would send a text message, one small chunk at a time, always looking at everything it has written so far to decide what comes next. That chunk is called a token, roughly a word or part of a word. The name just means “predicting from its own past output”.
This design is why models sometimes paint themselves into a corner. Once the first word of an answer is out, it is out, even if the reasoning that follows shows it was wrong. The model can only apologise and correct itself later, which is exactly what you sometimes see mid-answer. It is also why generation feels like typing rather than appearing all at once. Alternative approaches such as text diffusion produce whole blocks at a time and can revise them before you ever see them.