Every token the model produces is added to its own input for the next round. That feedback is what turns independent local guesses into a continuous piece of writing: each choice shrinks the set of continuations that still fit, so the text acquires a direction it was never given in advance.
Tracing a short sentence
Take the opening "The capital of France is". The next-token distribution is dominated by "Paris", so that token is chosen. The context is now "The capital of France is Paris". The next distribution is no longer about capitals; it is about what typically follows that completed statement, so options like a period, a comma, or a clause such as ", which sits on the Seine" become plausible. Each step is scored against a slightly longer history, and the sentence's shape is decided along the way rather than at the start.
Because later steps are conditioned on earlier ones, an early wrong turn tends to be carried forward rather than repaired. The model is not re-reading the whole answer and checking it against the world; it is continuing from whatever is already there.