Abstract:
Hateful online activity not only reflects underlying tensions but affects realworld behavior. Moderation policies should be based on existing laws to avoid new speech restrictions and dismiss claims of digital censorship. Germany's NetzDG regulation had instant online and offline effects, disrupting the channels that turn hatred into violence. Removing posts may distort online speech, but a "softening" of hateful expression through generative language models could reduce toxicity while minimizing distortions. Moderation rules must be clear and narrow to limit compliance costs without entrenching incumbents.