Specifically, do you think it violates US copyright law?

Companies like OpenAI and Stable Diffusion claim that their use of copyrighted material falls under Fair Use because it is transformative, provides many benefits to society, and would not be possible with only public domain training content.

Opponents claim that the scraping process does not fall under Fair Use because the purpose is commercial, creative, and harms the market for the original works. The training process likewise may be considered a copyright violation when the network weights end up containing the training data in a way that can be output verbatim.