What we need is quality control of training data, AI-generated or not. If most humans agree that picture X is "a dog", then that picture represents accurately t

What we need is quality control of training data, AI-generated or not. If most humans agree that picture X is "a dog", then that picture represents accurately that concept. The training datasets I've seen so far are awful and include human-generated garbage.