Show the remote candidate an AI's deficient answer to a well-asked question, and ask the candidate if they understand what exactly is wrong with the AI's assessment, or what the follow-up/rewritten prompt to the AI should be. Compile a library of such deficient chats with the AI.