Last year I commented[1] on the US court decision in Andersen et al. v. Stability AI Ltd., et. al. 23-cv-00201-WHO (N.D. Cal. Oct. 30, 2023), saying I believed the decision represented a significant triumph for the plaintiffs given the claims that were allowed to proceed.
The case involved a number of artists and illustrators suing various leading AI companies such as Stability, and Midjourney for using their copyrighted artistic work as training images for the defendants’ generative artificial intelligence (AI) algorithms without the plaintiffs’ permission.
In that decision, US District Judge William H. Orrick dismissed most of the plaintiffs’ claims (though I argued the most important claim of direct infringement survived intact,) but granted leave for the plaintiffs to amend their claims and address deficiencies the judge had identified.
The plaintiffs subsequently filed a First Amended Complaint (FAC), adding seven new artists as plaintiffs and a new defendant (Runway AI) to the case. All defendants moved to dismiss the FAC.
On 12 August 2024, Judge Orrick issued a decision on the motions to dismiss, allowing some of the plaintiffs’ claims to proceed to trial/summary judgment whilst dismissing some of the claims[2].
I believe the decision again represents a rather significant triumph for the plaintiffs given the claims that were allowed to proceed.
The followings are some key takeaways:-
1) Direct copyright infringement claims
The plaintiffs argue the Stable Diffusion model itself contains highly “compressed” copies of the training images, making the defendants’ use and distribution of Stable Diffusion a infringements. Judge Orrick was satisfied the plaintiffs addressed the deficiencies in its theory and had alleged additional facts as to how “the training images remain in and are used by Stable Diffusion”.
Runway, the newly added defendant, tried its hand in challenging this theory. The judge said whether or not the theory is supported by evidence is a matter to be decided later. The plaintiffs’ allegations are sufficient for the time being and the claim is allowed to proceed to summary judgment.
As discussed in my previous article, the direct infringement claims are the most important. They may also the ones the defendants will have the hardest time defending as they are completely independent of the output generated by the AI models. The fact these claims survive completely intact is a huge win for the plaintiffs.
2) Induced copyright infringement claims
The plaintiffs argue that by distributing Stable Diffusion for third parties to use, Stability is inducing third parties to infringe the plaintiffs copyright by causing others to copy their material.
Judge Orrick observed that “Stable Diffusion by operation by end users creates copyright infringement and was created to facilitate that infringement by design” is at least a “plausible inference”.
He also pointed to the plaintiffs’ contention the AI models can sometimes generate images identical or very similar to the training images, adding “[w]hether true and whether the result of a glitch (as Stability contends) or by design (plaintiffs’ contention) will be tested at a later date” and held the plaintiffs’ allegations of induced infringement are sufficient. The induced infringement claims are allowed to proceed.
These claims will likely rise and fall together with the direct infringement claims outlined above, though it must be noted that if AI companies are found to be inducing breaches by their users, their liability could be potentially limitless.
3) Copyright infringement claims against third party implementers of AI models
One of the defendants is the digital art website DeviantArt, which itself has not trained any AI model. It has only adopted AI tools provided (and trained by) others (such as Stability, a co-defendant) in its own AI image generation tool DreamUp (https://www.deviantart.com/dreamup), and consequently argues the plaintiffs’ direct infringement claims related to AI training cannot apply to it.
The Judge was satisfied plaintiffs had included additional allegations as to how their copyrighted work remains in Stable Diffusion and could be used in all Stable Diffusion versions (including in DeviantArt’s DreamUp) and held that “[t]he actual operation of Stable Diffusion 1.4 and whether the amount of any plaintiff’s copyrighted works in that program suffices for copyright infringement… remains to be tested”.
DeviantArt has warned that many other companies using the Stable Diffusion model (without taking part in its creation and training) could also be sued if the plaintiffs’ claims were not dismissed. It does appear its worst fears have come true.
4) False endorsement and trade dress claims
The plaintiffs allege their names appeared on a list of 4,700 artists posted by Midjourney’s CEO on its Discord server, where Midjourney’s CEO promoted the list as “describing the various styles of artistic works its AI product could produce”.
The plaintiffs also allege Midjourney has published user-created images that incorporate the plaintiffs’ names in its “showcase” site.
They argue the use of their names in this way constitutes false designation of origin (that the works generated by Midjourney originated from the plaintiffs) or false endorsement (that the plaintiffs endorsed Midjourney in any manner).
Midjourney counters the name list does not in itself support an inference of endorsement. The judge found that whether a prudent consumer would consider the name list and showcase as an endorsement is a matter to be tested at summary judgment.
The plaintiffs also argue Midjourney allows users to create works capturing “trade dress… [of the plaintiffs that] is inherently distinctive in look and feel as used in connection with their art…”.
Midjourney counters the plaintiffs have not adequately identified what constitutes their trade dresses and the plaintiffs are trying to monopolize unprotectable elements such as art styles.
The judge concluded Midjourney’s use of the plaintiffs’ names and its CLIP (Connecting text and images) model give rise to “plausible allegations that the CLIP model functions as a trade dress database and the use of its product to produce works based on the names of these plaintiffs”.
Whether such a claim can be substantiated is again a matter to be determined at summary judgment.
These claims are in essence unregistered trade mark infringement claims, highlighting other potential legal pitfalls AI companies may face apart from copyright claims.
5) Artists need not identify the works stolen, at least for now
Midjourney sought to have the copyright claims dismissed as the plaintiffs did not “identify specific, individual registered works that each artist contends Midjourney actually used for training”.
In a rather remarkable turn of events, the judge held: “Given the unique facts of this case – including the size of the LAION datasets and the nature of defendants’ products, including the added allegations disputing the transparency of the “open source” software at the heart of Stable Diffusion – that level of detail is not required for plaintiffs to state their claims. (emphasis added)”.
Even more remarkably, the judge further observed that “[the plaintiffs’ unregistered works] are relied on to support the plausibility of plaintiffs’ copyright theories (that all or most of the works in the LAION datasets were used by Midjourney and the other defendants to train their AI products, and that plaintiffs’ works or their protected elements that are contained in the AI products as the works or protected elements can be recreated by using the AI products). The FAC allegations and the exhibits help plaintiffs cross the plausibility threshold. (emphasis added)”.
This seems to go against (or at least blurs the line between) the orthodoxy that a plaintiff in a copyright infringement proceeding must specify which of their copyrighted work (and in the US, which registered copyrighted work) has been copied, seemingly lifting a formidable hurdle artists face in bringing proceedings against AI companies.
6) Fair use or not?
Perhaps not surprisingly, fair use makes an appearance in this case.
DeviantArt argues that any use of the plaintiffs’ work should be considered fair use and exempted from infringement liability.
It should be noted that any fair use argument would inevitably involve detailed examination of evidence as to how the copyrighted work has been used by the defendants, which the Court should not go into in dismissal proceedings. Judge Orrick (rightly in my opinion) pointed this out in rejecting DeviantArt’s motion to dismiss.
Fair use is THE main defence of AI companies in the US.
AI companies argue that using copyrighted work to train AI algorithms is no different from an art student learning how to paint by copying/painting existing masterworks.
AI companies point to the landmark case of Authors Guild v. Google, Inc. 804 F.3d 202 (2nd Cir. 2015), where the Second Circuit Court held that “Google’s unauthorized digitizing of copyright-protected works, creation of a search functionality, and display of snippets from those works are non-infringing fair uses. The purpose of the copying is highly transformative, the public display of text is limited, and the revelations do not provide a significant market substitute for the protected aspects of the originals. Google’s commercial nature and profit motivation do not justify denial of fair use.”
AI companies argue their AI modules aim to transform existing works to create new ones and any use of copyrighted work in training for this transformative process is justifiable.
The 2023 US Supreme Court case of Andy Warhol Foundation for the Visual Arts v Goldsmith 598U.S. 508 (2023) may have thrown a spanner into the fair use defence.
In that case, the majority of the Supreme Court stressed that assessment of fair use should be based on the commercial nature of the use and whether the derivative work represents direct commercial competition to the original work.
Fair use is unlikely to be found when an original work and its derivative (even if the derivative is transformative) share the “same or highly similar purpose” and that use of the derivative work is commercial in nature.
With widespread reports of AI-generated work displacing and replacing human-created ones and artists potentially being put out of work by AI, the fair use defence by AI companies may have suddenly become an uphill battle.
7) Rough waters ahead for AI companies?
The copyright infringement case now proceeds to discovery, where both sides are obliged to disclose all information and documents relevant to the proceedings.
AI companies may face the prospect of being forced to disclose uncomfortable information, such as how they harvested training data; what deals they made to obtain the training data; the nature of their relationship with database operators; how training data was actually used in the training process.
Legal troubles aside, these disclosures may have far-reaching public relations and political ramifications, potentially forcing governments to create new regulations.
No matter what, this saga is far from over and a number of fierce battles lie ahead.
For further information, please contact:
Anthony Leung, Haldanes
anthony.leung@haldanes.com
[1] https://www.haldanes.com/publications/artists-rage-against-the-machine/
[2] The Plaintiffs’ DMCA (Digital Millennium Copyright Act) claims, breach of contract claims were dismissed with prejudice (i.e. the same claims cannot be raised again); the Plaintiffs’ unjust enrichment claims were dismissed with leave to amend.
