SAN FRANCISCO (CN) — A set of companies that offer generative artificial intelligence image software argued Wednesday afternoon for the dismissal a direct copyright infringement claim from artists who claim the companies used their work without permission to train a text-to-image AI model.
According to the artists’ their class action, AI software company Midjourney and online art platform DeviantArt used their art to train text-to-image AI tool Stable Diffusion without paying the artists or getting consent to use their work as training images, and the resulting images continue to compete against the human artists' work in the marketplace.
The artists first sued DeviantArt, Midjourney and Stable Diffusion creator Stability AI last year, but U.S. District Senior Judge William Orrick tossed most of the artists’ claims in October 2023, allowing only a single direct copyright claim to survive against Stability AI.
An amended complaint from the plaintiffs prompted a tentative ruling from Orrick on Monday where he said he intended to allow direct and induced infringement claims against all the defendants to proceed to discovery because the artists have plausibly claimed that copies of their work are stored in versions of Stable Diffusion.
The artists' Digital Millennium Copyright Act claims, however, should be tossed, according to Orrick, because the plaintiffs have not plausibly pleaded that the defendants reproduced or distributed their works.
Andrew Gass, counsel for DeviantArt, argued Wednesday that the art platform should be dismissed from the action because it merely used Stable Diffusion to produce images for its DreamUp AI tool but has never directly used the artists’ images to train an AI model or used an AI model to create images that look like the artists’ images, or used the technology in any other relevant way.
“There is no allegation that DeviantArt itself ever extracted from the model any output that is a copyright infringement of the plaintiffs works. There's no allegation that any of DeviantArt users have ever extracted any output from DeviantArt’s implementation of the model,” Gass said.
Gass said all DeviantArt did was download the model and make it available — they did not scrape data or train anybody or anything on it. He characterized it as “classic fair use.”
“That leaves us with the idea that an actor situated the way DeviantArt is is somehow liable for copyright infringement for the vestigial ghost copies that are in the model somehow, that they never accessed,” Gass said.
He told Orrick that there are thousands of other users like DeviantArt that downloaded the Stable Diffusion model or made it available, and that they should not be dragged into discovery unless the plaintiffs could plead “something beyond an intermediate invisible copy that no one ever accessed.”
Laura Matson, counsel for the artists, replied that they have evidence that a researcher at DeviantArt generated distinct models using the artists’ work.
“The reason that DeviantArt is liable for infringement is because it has copied statutory copies of protected work,” she said.
Orrick asked Matson to name the specific harm that her plaintiffs suffered from DeviantArt’s claimed actions.
“What the plaintiffs are suffering from is that there are copies of their copyright protected work that are contained within this model. They are compressed, they're contained within this model, and we have alleged that they're reproducible,” Matson replied.
As for fair use, Matson said it was inappropriate to discuss it at the motion to dismiss stage, but said that the plaintiffs were harmed in the marketplace because of DeviantArt’s conduct, which would undercut the fair use defense.
Midjourney had much of the same defense as DeviantArt.
“Plaintiffs have yet to be able to show a single example of a work they’ve extracted from Midjourney through prompting that looks like their copyright registered work,” Angela Dunning, Midjourney's counsel, said.
Orrick asked Christopher Young, also representing the artists, why he should not follow U.S. District Judge Vincent Chhabria’s precedent set in a similar case where comedian and author, Sara Silverman, along with other authors, sued Meta over its use of their books to train a language model.
In that case, Chhabria ruled that the claims should be dismissed because a model cannot infringe copyrighted works unless it reconstitutes, reuses or readapts the book.
Young replied that they have presented more facts in this case than the Silverman case, and that image generators are fundamentally different from text generators, so the two cases should not be compared.
Orrick took the matter under submission.
Subscribe to Closing Arguments
Sign up for new weekly newsletter Closing Arguments to get the latest about ongoing trials, major litigation and hot cases and rulings in courthouses around the U.S. and the world.
