You're right, sorry, I had several tabs open on a similar subject... the post I was referring to is this:

[N] Class-action law­suit filed against Sta­bil­ity AI, DeviantArt, and Mid­journey for using the text-to-image AI Sta­ble Dif­fu­sion : MachineLearning

> The multi-part fair use test established in AGI vs Google is widely held to be applicable to AI and ML models.

The US four-part fair use test was established long before AGI v. Google: in the 19th century in Folsom v. Marsh. It was encoded into copyright law in 1976. It's applicable to everything.

The case only decided that Google's specific book service did in fact pass the test. The most important aspect is that the judge found that there was no economic damage to the book authors, that it did not replace the books or negatively impact the market for books.

The decision is not applicable to other projects that may be substantially different in character. I'm sure OpenAI's lawyers are hoping that DALL-E will be considered to be equivalent to Google's book search - that they have fair use rights to digitize copyrighted material without permission, and publish something transformed that only contains "snippets" of it. But they will have to get around the fourth factor. Who will commission an expensive original artwork from Greg Rutkowski, when they can simply type a prompt including "in the style of..." and get something substantially similar, for less than a nickel? Will companies use GPL3 code in their products, when they can get a mashed-up facsimile with the restrictive license removed? The question of fair use in the context of generative neural networks is far from settled; hence the lawsuits in the (other) post.