OpenAI archives every request and output text. Why not compare suspected A.I. text against this?