How it changes a reply
As an AI model writes, it chooses each next token, or piece of text. At a low temperature, it strongly favours the most likely choices, so repeated requests often get similar replies. At a high temperature, less likely choices become more common, making replies more varied but sometimes less coherent.
What it does not control
Temperature does not measure truth or creativity. A low setting can give the same wrong answer repeatedly, while a high setting can still give a correct one. Even at the lowest temperature, some services may not produce identical replies every time.