A research paper from the 2021 top cryptography conference CCS demonstrated that even with a classifier accuracy of 99.9%, scanning for low-frequency illegal content still produces a massive number of false positives.