> It's proved that they just take a brunch of poor peoples poorly paid to deal with real messages and try to moderate them sometimes with a poor understanding of the context, and use it as an input to teach the IA how to do "proper" moderation.

That's exactly what the article says: "you’d write up some guidelines (...) and then contract with some external company to have human beings read those guidelines and rate lots of examples that you send them" and then it mentions people from places with low salaries, like rural India and the Philippines.

And then in the possible solutions, it mentions hiring people well steeped into the specific cultural context of those you're trying to moderate.

This is where I think the article undermines itself, because if paying for some different (and probably more expensive) set of people may be a solution, then the question that was supposed to be answered remains: "why do companies with unbounded resources not do that?"

While hiring those people might be more expensive, I expect it isn't much. Annotators willing to read a set of guidelines and perform rating tasks long term are generally not cheap. I suspect that for something like twitter, identifying people most suited to moderate a particular piece of content is itself a very difficult technical and social problem. So even assuming you can figure out how to find these people and hire them, there's still a long path from there to improving moderation.