I'll bet adversarial attacks are possible - ones where you can add some innocuous text to a paper (or alter the existing text in seemingly minor ways) and significantly increase the likelihood of acceptance by some given model