"I'd expect to see particular themes and unique writing styles emerge repeatedly over time from a human author. I'm not sure if we'd see that from an AI"

On the other hand, a bulk of submissions by a natural language AI might have an all-too-obvious tell; perhaps trending artifacts, including themes that act as a giveaway. I wonder if: 1) a sort of anti-aliasing would be a useful addition to the algorithms producing papers (if not already a feature), and 2) if that would serve as a successful way to mitigate the risk of recognition, if indeed as per your example a professor would be likely to grade less on par with human submissions for AI papers.