I feel like this approach might be relevant to achieve the equivalent of “go sit in the corner and think about what you’ve done.” Pull out two random poorly-rated chatbot conversations and have the model look for commonalities in its shortcomings.