I would think it unlikely that they have a comprehensive corpus of "offensive content", meaning they can't train/test their system on all of it. So really they'

I would think it unlikely that they have a comprehensive corpus of "offensive content", meaning they can't train/test their system on all of it. So really they're claiming "99.995% accuracy against what we already have". The real world experience will surely be different. You're right that the number is essentially meaningless.