AI companies work around this by paying human classifiers in poor but English-speaking countries to generate new data. Since the classifiers are poor but not stupid, they augment their low piecework income by using other AIs to do the training for them. See, AIs can save workers labor after all.

On the one hand, you'd think the AI companies would check to make sure they aren't using AI themselves and damaging their models.

On the other hand, AI companies are being run by some of the laziest and stupidest people alive, and are probably just rubber-stamping everything no matter how blatantly garbage the results are.