As dumb as twitch moderation can be, the issue is that you can't possibly task twitch moderation with knowing to which degree an offensive joke is relevant or not. If an AI is set up to spout something-phobic stuff 24/7 surely they are going to ban that, rather than wondering where is the degree of offensiveness that warrants a ban, they just ban when ToS is breached.

At least that's what would make sense, and Twitch moderation is far, FAR from making sense in general, but in this one thing they might be reasonable.