I would take the claims of ethics concerns more seriously if the training set data was more ethically sourced. I buy that the law probably[0] considers scraping

I would take the claims of ethics concerns more seriously if the training set data was more ethically sourced. I buy that the law probably[0] considers scraping the Internet to train AI legal, but there are various non-copyright concerns from this approach. Such as GPT-3 remembering people's phone numbers or DALL-E remembering a particular person's X-rays or CAT scans. And even if you buy the copyright argument, it does nothing for the users of these systems who are now downstream of an unbounded number of copyright claims in the future. [0] AI companies are banking on Authors Guild v. Google being controlling precedent in the US. EU law already explicitly allows training AI on copyrighted data.