Actually there isn't much "control", beyond the awareness that X model is trained on a dataset scraped from Y, and basic cleaning/sanitizing. There's so much da

Actually there isn't much "control", beyond the awareness that X model is trained on a dataset scraped from Y, and basic cleaning/sanitizing. There's so much data in use that it'd take decades for a human team to curate or generate in a way that meaningfully balances the datasets. And so models are also used in curation and generation, which themselves are blackboxes...