Stability AI trained its image generator on the LAION-5B data set of 5 billion (!!!) image-text pairs. If you download this data set, you don’t get the images, you get metadata about each image including its URL, plus the nearby text and caption. You then feed your beefy computer the metadata and it retrieves the image at the URL and uses it to train the Runway ML stable diffusion algorithm (Stability AI has done a good job obscuring how little novel work it did). Your computer doesn’t permanently store the images, the neural net studies them like an art student. So if this training process is illegal and requires permission, then every search engine and web crawler in the world is illegal, and any artist who visits a museum or borrows an art book from the library should be quizzed “Hold on,, you look like you might be training on that copyrighted image!”
I feel for artists whose work is being devalued by image generators that can effortlessly make “large painting of child gazing up at two broken robots in a field by Simon Stalenhåg” (see images of some of his actual paintings). It feels wrong, it feels like the artist deserves something. But feels aren’t law. The digital image is not a painting, and unless someone commits fraud by claiming it’s actually by the artist, nothing illegal has happened in most jurisdictions. Even if there was some way for the artist to get a cut (maybe artists will start claiming design patents on their style), 20% of $0 is nothing.
AI-generated art is mostly derivative art… but all art is “derivative art”. “Should” isn’t law. If someone tries to pass off an AI-generated artwork as being by a particular artist, they are committing fraud. Until that occurs, making an image in the style of an artist is obviously incontrovertibly legal. So too was the training process where the AI “looked” at billions of image-text pairs on the web and developed an internal representation of all art approaches and most famous artists’ styles; it’s merely a far more diligent student of art than any human being. It feels like artist should get something, so please say how the law should change to make that happen. But if the artist gets a percentage of the value of the image ripping off their work, and the image is uncopyrightable AI slop, then 20% of $0 is zero.
You’re welcome to feed my wise words into a video generator to produce a Two Minute Paper on the subject 🙂.
Actual judges weigh in
The lawsuits from aggrieved artists are wending their way through the courts. A big one against Anthropic was ruled on by Judge William Alsup in 2025, legendary for (allegedly) learning to program so he could better rule on Oracle’s lawsuit against Google’s implementation of Java. National Law Review reported on the two parts of this. Judge Alsop ‘ruled that AI company Anthropic’s use of copyrighted books to train its powerful AI model, Claude, was “exceedingly transformative” and qualified as a legal “fair use. …’
So the claim that AI companies are ripping off creative artists by training on their work is on very shaky legal ground unless other judges ignore Alsop’s conclusions or Congress changes the law.
But flat-out piracy…
National Law Review continued ‘However, the judge drew a sharp line in the sand, ruling that Anthropic’s separate act of downloading and storing millions of books from “pirate sites” was not a fair use and that the company will have to face a trial for it.’
Indeed, the shortcut that nearly all AI companies took of downloading from pirate sites instead of purchasing a copy to train on will deservedly come back to bite them. Judge Alsup ruled: “This order doubts that any accused infringer could ever meet its burden of explaining why downloading source copies from pirate sites that it could have purchased or otherwise accessed lawfully was itself reasonably necessary to any subsequent fair use.” “That Anthropic later bought a copy of a book it earlier stole off the internet will not absolve it of liability for the theft, but it may affect the extent of statutory damages.”
7 million books × $150,000 maximum statutory damages is ONE TREEELYUN DOLARS; Anthropic and the Authors Guild later settled on a deal to pay authors of only 482,000 books that Anthropic illegally pirated $3,000 for each book; Alsup criticized the deal but it’s gone through; AP reports “About 91% of the more than 482,000 books covered by the ruling have been claimed by authors or publishers who are now due payment.”
On the other hand…
Later in 2025 Judge Vince Chhabria ruled on a similar case. Whitecase reports that he also found training “highly transformative”. But Judge Chhabria said Alsup improperly focused on the transformative nature of training to the exclusion of potential market harm, “blowing off the most important factor in the fair use analysis.” In other words if artists can show that AI harms or dilutes the market for their works, it may not be fair use. Judge Chhabria also noted that “the consequences of this ruling are limited” because “these plaintiffs made the wrong arguments and failed to develop a record in support of the right one.”
