Hey Florian, thanks for good questions. Let say one option is what ChatGPT did. Let the users to like or dislike the answers what can adjust the weight. But this also bring the issue that someone will purposely do the opposite. Or that the user without domain knowledge is satisfied with answer what is wrong and approves the answer. For general purpose AI's, i think the answers will be more precise for general questions as there is good data set. But what is with domain specific questions. For that you need training data specified around that domain provided by some credible source. I saw that "gradually reasoning" helped the general purpose AI to give good answers even for domain specific questions.

It will be interesting when "Big Tech" companies release more general purpose data set what can be retrained around some specific domain.