Perhaps I should say, deterministic behavior to satisfy pedants?
E.g. "deterministic?" A naive Bayes model that optimizes for inflammatory text keywords is deterministic. Is it what you want?
I think you're making the point that consumer comprehension of how a system works is important, but at what level is that comprehension needed?
When you use email, you know what's going to happen. You can predict the behavior (yes, there are spam filters).
Facebook (and other social media) behave such that if we knew the entire sequence of user behavior, we wouldn't be able to predict what happens next. It's this sort of black box behavior that isn't good.
If you have an algorithm that is strictly optimizing to reduce inflammatory text, then I suppose that's ok. But still, it might be better to have user-moderation instead. I'd rather people just be banned by moderators.