> simply acquiring only training data you have permission to use
Currently it's generally infeasible to obtain licenses at the required scale.
When attempting to develop a model that can describe photos for visually impaired users, I had even tried to reach out to obtain a license from Getty. They repeatedly told me that they don't license images for machine learning[0].
I think it's easy to say "well too bad, it doesn't deserve to exist" if you're just thinking about DALL-E 3, but there's a huge number of positive and far less-controversial applications of machine learning that benefit from web-scale pretraining and foundation models - spam filtering, tumour segmentation, voice transcription, language translation, defect detection, etc.
