A Mastodon user stumbling upon one of these comments could easily assume that it is just another fully independent “toot” (Mastodon’s equivalent of tweet).

Wait, back up… Mastodon calls these “toots”? So, everybody is posting farts?

That’s the thing about automation and training models.

First, they implement some sort of auto-reporting bot that requires a human to review them. In the beginning, it only about 50% accurate, but as they give it more and more examples of good and bad results through the human reviews, it moves to 80%, then 90%, then 99%, then 99.99% accuracy.

After a while, the humans on the other end are so numb to the 9999 entries they have to mark as approved that they can barely tell what’s a rejection themselves, and the moderation team is asking itself just what this human review is actually doing. If it’s 99.99% accurate, why not let the bot decide?

Then, the model moves on from auto-reporting to auto-moderation.

but is concerned about hosting fees for serving images to millions of people

People stopped caring about image bandwidth decades ago. Try wrangling a video-hosting problem, like PeerTube does.