[Michelangelo11]

You can't fully automate AI content for a number of reasons, chiefly because it can generate something extremely insensitive or offensive that can torpedo the whole business (e.g. AI Seinfeld's transphobic joke), but also because gettibg AI output to stay within a certain range of acceptable responses is -incredibly- hard, and if the ratio of misses to hits is too high, you're losing money (as every shot, whether you hit or miss, incurs some costs). For more details about this, look up how AutoGPT and similar things are actually doing in practice -- they get stuck a lot and require constant supervision.

And this is because, again, keeping AI outcomes within a certain range at scale is both amazingly hard and computationally irreducible (i.e. you can't predict its responses at a scale of say 1 million shots, without taking 1 million shots).

[BizarroLand]

Could you not make a censor AI that scans content for offensive metadata and rejects the ones it finds?

There may still be a bit of slipthrough but it should be dramatically mitigated.