“Currently, humans evaluating responses from models like ChatGPT or Claude are not yet considered "weak supervisors" because they can generally determine the quality of the output.”