“OpenAI uses an automated procedure involving GPT-4 to judge the factuality of long-form answers by checking for consistency with a human-checked reference answer.”