The point is that this leads to an arms race. If Arxiv uses a top-of-line LLM for, say, 20 minutes per paper, cheating authors will use a top-of-line LLM for 21 minutes to beat that.