Breaking: Judge Orrick dismissed most of the claims brought by the artists in Andersen v. Stability AI.
This was an unsurprising result. Just because the technology is "new" doesn't mean we disregard current law. The claims were doomed regardless of AI.🧵acrobat.adobe.com/id/urn:aaid:sc…
The holding isn't precedential. It reaffirms long standing copyright principles:
(1) General pleading is not enough. Plaintiffs must identify specific works that were allegedly infringed.
(2) No infringement for outputs that are not substantially similar to a protected work.
Quick recap:
Andersen et al allege Stability AI infringes works by providing the works to Stable Diffusion for training. They also allege all Stable Diffusion outputs are derivative because the training data consists of protected works.
Plaintiffs fail to identify any works.
Plaintiffs also allege that DeviantArt and Midjourney are implicated because they rely on Stable Diffusion's library for their AI services too.
(1) Registration
Plaintiffs Mckernan and Ortiz failed to register their works with USCO so no relief is available under the Copyright Act. Their claims are dismissed w/prejudice.
Andersen is limited to the works she registered.
2. Identifying works
Andersen's claims regarding her works being copied into the training sets can proceed to discovery. Though she doesn't identify any specific works, her inference based on the output pages displayed on is plausible. haveibeentrained.com
3. Direct Infringement (Stability AI)
Stability AI's motion to dismiss is denied. Court argues that whether Stability AI's copying for its training data cannot be decided at this stage.
This is unsurprising. The copying Q (including the fair use analysis) is fact intensive.
Though keep in mind, old Google Search case law foreshadows the inevitable. Field v. Google dealt with similar facts / issues: does Google infringe when it scrapes publicly available data for indexing?
No -- indexing is a transformative use. The same can be said here.
4. Direct Infringement (DeviantArt).
The Q is whether DeviantArt directly infringes when it provides their users access to the Stable Diffusion library.
It's less likely if Diffusion just contains instructions for content creation. Plaintiffs need to provide more facts.
BTW, the Plaintiffs managed to shoot themselves in the foot by stating that the output images can't possibly be similar to the allegedly "compressed" versions of their works in the Stability training sets, arguing that derivative should just be assumed.
The Court doesn't buy it.
Notably the Court also doesn't buy the Plaintiff's "compressed files" theory (i.e. that Stability AI takes protected works and merely compresses those works for the training sets).
Stability AI vehemently rejected this allegation. The Judge tells Plaintiffs: prove it.
Court also notes that it's "simply not plausible" that every image used to train Stable Diffusion was protected, and thus not plausible that all output images are derivative.
In other words: Plaintiff's lazy pleading tactics rightly fail. Substantial similarity requires works.
Cite to Google Books: "a work is not derivative unless it has been substantially copied from the prior works."
The Court grants DeviantArt's motion to dismiss.
BUT the Court also generously provides Plaintiffs leave to amend their DeviantArt complaint to identify specific facts and works in the training data / outputs that substantially infringe their own works.
(my guess is Plaintiffs simply do not have such facts to plead).
5. Direct Infringement (Midjourney)
Again the Court asks: where are the facts. Plaintiffs didn't even plead how Midjourney goes about training its models.
Plaintiffs only allege that Midjourney uses Stable Diffusion. The Court asks: but how?
Because Plaintiffs failed to properly plead direct Infringement for DeviantArt and Midjourney, the vicarious liability claims are dismissed (w/leave to amend). Unsurprising.
(Stability AI) Similarly fails -- Judge demands facts re: Plaintiff's "compressed works" theory. Even then, Judge remains dubious: "Plaintiff's complaint is devoid of any allegations that [the artist's works] were used to create "fakes" of their work."
Dismissed / leave to amend
7. DMCA (removal of copyright management info)
More facts needed. Plaintiffs never identified which of their online works contained CMI. BECAUSE THEY NEVER ALLEGED ANY SPECIFIC WORKS TO BEGIN WITH 🤷🏻♀️
Plaintiffs also failed to allege which Defendant actually stripped the info.
8. Publicity Rights
You guessed it -- no facts. Plaintiffs first alleged that the defendants infringed their artistic styles (preempted by copyright law). They try again, alleging the defendants improperly used their names to advertise the product.
The Court says: where??
9. DeviantArt's 1A argument
Court says DeviantArt can re-raise their transformative use defense once the Plaintiffs get their facts straight.
10. Unfair Competition Law
"Given the shifting nature of plaintiffs’ claims – from misappropriation of their styles or artistic identities to misappropriation of their names and associating their names with works not their own – the common law UCL claim is likewise dismissed."
11. Breach of Contract
Plaintiffs allege that DeviantArt broke their own TOS by incorporating the Stable Diffusion library.
Court: no facts provided to support this allegation nor that Plaintiffs are beneficiaries entitled to enforce such contract between Stability/DeviantArt.
Overall good result.
Where there is money to be made from technological advancements that facilitate content creation, there will always be rightsholders lurking around the corner to cash in.
The Andersen case is yet another rightsholder grift. There will be more.
@threadreaderapp unroll.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
Yesterday, @ProgressChamber submitted comments responding to the U.S. Copyright Office's Notice of Inquiry on AI and Copyright.
In sum, we suggest that existing copyright law and fair use principles adequately address the latest advancements in #GenAI.🧵 acrobat.adobe.com/id/urn:aaid:sc…
1. The capabilities of Generative AI can be a foundation for idea formulation and inspiration for artists and creators more generally. Among other things, Gen AI improves content moderation, revolutionizes medical research, enhances education, and bolsters autonomous vehicles.
Indeed, the societal benefits of Gen AI are readily apparent. Policymakers must keep these benefits in mind as they approach regulation. Otherwise, we may never realize the tech's true potential.
Copyright law is one major threat to AI that could deliver that crushing blow.
Note -- the complaint is heavily redacted throughout, making it difficult to opine on some of the claims (such as COPPA). Those facts matter, so we'll have to wait for more details to come out.
The dissent finds the majority's decision to stay the injunction "highly disturbing" noting that the case is unlikely to be resolved until spring of next year, allowing government to interfere with social media's editorial decisions until then.
The dissent also argues that the potential harms raised by the govt (such as precluding govt actors from speaking publicly about sensitive events) are unfounded and speculative, undeserving of the stay.
The CA case involves numerous complaints by minors alleging addiction claims. The issues raised here are similar if not identical to the issues raised in the federal school district MDL (ongoing). Same analysis follows.
Social media companies are not products for the purposes of products liability law. Court instead proceeds on the negligence claims, similar to the ones arising out of Snapchat's speed filter in Lemmon v. Snap.
Social media companies are not products for the purposes of products liability law. Court instead proceeds on the negligence claims, similar to the ones arising out of Snapchat's speed filter in Lemmon v. Snap.
While the govt may have a limited right to restrict the manner of speech in order to protect unwilling viewers of the public, it is expressly forbidden from restricting willing adults accessing legally protected speech.
The latter is the essence of school district suits.
The Court emphasizes Act 689's failure to reach sites like Parler, Gab, and Truth Social (a recurring problem).
If the intent is truly to protect kids from awful content, why not include the sites responsible for some of the most heinous and hateful content produced online?