Major test shows AI hate-speech filters disagree widely, creating uneven standards across groups and platforms.