One of my extended family members works for a FAANG company (not Facebook) on a team that moderates user-generated content. They are actually using AI to great effect, with great results. They have large amounts of manual review and moderation as well, but having AI augment the human intervention and also continuously trained based on human overrides of AI decisions makes the system much more efficient.
What many people don't understand is that even if Facebook employed human moderators to individually review every group, there would still be isolated examples of policy violations slipping through. Human moderators are far from perfect. They can and do make mistakes. At scale, you get some moderators who don't actually care about doing the work, so they start letting things slip through instead of reviewing them. AI is also helpful in flagging human moderator decisions that disagree strongly with AI predictions, which can then be used to catch moderator errors or improve the AI, depending on which is ultimately deemed correct.
It's easy to forget that Facebook has almost 3 billion users. At scales that large, even 99.99% correct content moderation (human or otherwise) will still result in a lot of incorrectly moderated content slipping through.
More commonly, it may not be immediately obvious that content breaks policies, or in this case that a group is political. Or maybe the group started as one thing, but then evolved over time to become highly political. Or maybe they chose a name that sounds innocuous, but is actually very offensive given some obscure context or lingo. The problem is impossible to solve perfectly, so we need to instead focus on setting realistic expectations for what can be done.