I recognize that ignoring human costs is incredibly convenient, but I wish more people could really internalize that the only thing keeping LLMs from routinely spitting out the monstrous garbage in their training data is the armies of underpaid human moderators, usually in developing countries, seeing and filtering out the worst things in the world for pennies an hour. LLMs rely on classic extractive-resource colonialism, but the resource being extracted is psychological health and safety.

@mhoye @gvwilson do you have a concurrent citation for this? Certainly seems consistent with what I know about the players involved but i haven’t seen reporting on this specifically yet