It doesn't involve someone explicitly telling them what they got wrong or right. They use something like supervised learning with a large amount of data. You can do something after that called “fine tuning” where you provide example responses, but I think you’re asking how the LLMs “know” something. And afaik it mostly just tries to guess the next word and does such a good job after being trained that it APPEARS to know things. 🤷‍♂️