Introduction
The question as to whether the generative AI (GenAI) systems and their use infringes copyright is a hotly debated one, with no fewer than 12 proposed class actions and 3 lawsuits in the United States and one in the United Kingdom already focused on resolving this question. The question is a central one for the creative community as well as for the GenAI platforms and, as the recent suit involving the New York Times against Open AI and Microsoft shows, it is permeated with significant legal and policy issues. it is of, course, and very interesting issue for copyright lawyers including copyright lawyers in Canada.
So, far while there has been lots of litigation, and four court decisions, there has been no definitive decisions on these questions. However, the four interlocutory decisions issued to date provide some insights into how courts are approaching the infringement claims.
This blog post focuses on these four court decisions:
- Andersen v. Stability AI LTD., (N.D. Cal. Oct. 30, 2023)
- Kadrey v Meta Platforms Inc., (N.D. Cal. Nov. 20 2023)
- Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc. (D.Del.Sep. 25, 2023)
- Getty Images (US) Inc & Ors v Stability AI Ltd [2023] EWHC 3090 (Ch) (01 December 2023)
The questions canvassed in these decisions include these copyright infringement questions:
- Whether downloading and copying of works (or other copyright subject matter) to train GenAI models infringes copyright.
- Whether the trained GenAI models are infringing because they contain or embody copies of training materials in a material object or material form or are, under US law, derivative works.
- Whether all outputs from GenAI systems are derivative works because they are derived from copyright works and whether specific outputs can infringe the copyrights in specific training materials.
- Whether outputs generated from an GenAI system located in the U.S. but accessed by users outside the US can infringe the making available right.
- Whether offering a GenAI system for download from the U.S. to users outside the U.S. such as in the United Kingdom could constitute secondary infringement under the law of the country of access.
- Whether in the fair use is potentially available in the U.S. as a defense to GenAI copyright infringement claims.
My key takeaways from these cases is provided after my summaries of them below.
Note of caution
Facts always matter. The decisions issued to date are all based on disputed facts, none of which have been authoritatively ruled on by the courts. They are also premised on different GenAI systems which were trained, and offered to the public, in different ways. The cases to date were also either decided based on pleading or summary judgment motions where the courts did not have sufficient facts to make binding determinations either as to the facts or the law. They are also based on the specific GenAI systems that are the subject of the complaints and on the specific claims, defenses, and arguments advanced by the parties in these cases. The decisions are also based on the applicable local law, i.e., U.S. and UK law and therefore may not be applicable or persuasive under other laws. Therefore, while the decisions provide insight into the arguments for and against the infringement and defense theories, and show how courts have considered various arguments, they cannot be taken as binding decisions on the merits, such as decisions rendered after a full trial and after all appeals have been exhausted.
Another consideration is that the liability theories against defendants have included many claims besides copyright infringement. These include infringement claims related to dealings with rights management information (such as violations of the DMCA), contract and trademarks related claims, and various U.S. State law claims including for breaches of publicity rights and tort claims such as negligence. This blog focuses only on copyright infringement claims with a specific focus on direct infringement and (UK) secondary liability claims and defenses such as U.S. fair use.
One further caveat. This post does not express my opinion on the issues canvassed in the decisions. Rather, the attempt here is only to summarize the decisions and to help unpack what the courts have said to date.
Andersen v. Stability AI
This U.S. decision involves a putative class action of artists that challenge the creation and use of Stable Diffusion, a GenAI software product. The plaintiffs allege that Stable Diffusion was “trained” on plaintiffs’ works of art to be able to produce output images “in the style” of particular artists. The three sets of defendants are (i) Stability AI Ltd. and Stability AI, Inc. (“Stability”); (ii) DeviantArt, Inc.; and (iii) Midjourney, Inc..
Each of the defendants filed separate pleadings motions to dismiss claims. The court largely granted the defendants’ motions to strike, with leave to amend the Complaint to provide further clarity regarding the alleged liability theories against each defendant.
BACKGROUND
In Anderson v Stability AI, the plaintiffs allege that Stability created and released a “general-purpose” software program called Stable Diffusion under a “permission open-source license.” Stability is alleged to have “downloaded of otherwise acquired copies of billions of copyrighted images without permission to create Stable Diffusion,” known as “training images”. Over five billion images were allegedly scraped (and thereby copied) from the internet for training purposes through the services of an organization (LAION, Large-Scale Artificial Intelligence Open Network) paid for by Stability.
Stable Diffusion is alleged to be a “software library” providing “image-generating services” for products produced and maintained by the defendants including DreamStudio, DreamUp, and Midjourney. According to the decision, users make use of these products by entering text prompts into the programs to create images “in the style” of artists. The new images are created allegedly “through a mathematical process” that are based on the training images and are alleged to be “derivative” of the training images.
DreamStudio is Stability’s product. It functions as a user interface accessing a trained version of Stable Diffusion. Deviant Art’s product, DreamUp, is a commercial product that relies on Stable Diffusion to produce images. The plaintiffs allege that at least one LAION dataset that was incorporated into Stable Diffusion for training images was also scraped from primarily 100 websites and used to train that product.
Midjourney is alleged to be a commercial product that produces images in response to text prompts in the same manner as DreamStudio and DreamUp. The plaintiffs allege that Midjourney uses Stable Diffusion but also that it was “trained on a subset of the images used to train Stable Diffusion.”
While the plaintiffs assert a number of claims against the defendants, I focus here only on the claims for direct copyright infringement.[i] Note, the applicability of the fair use defense was not canvassed in the decision.
Direct Infringement Allegations Against Stability AI
The Plaintiffs’ primary theory of direct copyright infringement is based on Stability’s creation and use of training images scraped from the internet into the LAION datasets and then used to train Stable Diffusion. The two claims, which the court refused to strike out, were the claims that the downloading and copying of images made for training Stability Diffusion and the storing and incorporation of the images as “compressed copies” within Stable Diffusion were properly pleaded causes of action.[ii]
Direct Infringement Allegations Against DeviantArt
The plaintiffs advanced three claims of direct copyright infringement against DeviantArt’s DreamUP GenAI product.
(1) direct infringement by distributing Stable Diffusion, which contains compressed copies of the training images, as part of DeviantArt’s DreamUp AI imaging product; (2) direct infringement by creating and distributing their DreamUp, which is itself an infringing derivative work; and (3) generating and distributing output images which are infringing derivative works.
Whether Stable Diffusion contains infringing copies of Training Images
DeviantArt disputes the assertions that “embedded and stored compressed copies of the training images” are contained within Stable Diffusion. DeviantArt (and Stability [notwithstanding the finding above related to direct infringement against Stability] and Midjourney) argue that those assertions are implausible given plaintiffs’ allegation that the training dataset was comprised of five billion images. They argue that “five billion images could not possibly be compressed into an active program.” They also argue that the plaintiffs made contradictory admissions in the Complaint suggesting that the training of Stability Diffusion involved not copying of images, but instead the application of mathematical equations and algorithms to capture concepts from the Training Images.
The court found that the Plaintiffs’ Complaint was unclear. The court appears to accept that a claim that a GenAI product may contain or incorporate infringing copies (referred to as “compressed copies”) may make the product infringing. However, the court was not satisfied that the claim related to “compressed copies” was pleaded with sufficient clarity.
Turning to the first theory of direct copyright infringement and the plausibility of plaintiffs’ assertion that Stable Diffusion contains “compressed copies” of the Training Images and DeviantArt’s DreamUp product utilizes those compress copies, DeviantArt is correct that the Complaint is unclear. As noted above, the Complaint repeatedly alleges that Stable Diffusion contains compressed copies of registered works. But the Complaint also describes the diffusion practice as follows:
Because a trained diffusion model can produce a copy of any of its Training Images—which could number in the billions—the diffusion model can be considered an alternative way of storing a copy of those images. In essence, it’s similar to having a directory on your computer of billions of JPEG image files. But the diffusion model uses statistical and mathematical methods to store these images in an even more efficient and compressed manner.
The decision did not directly engage with the U.S. Copyright Act definition of “Copies” to connect the comment that “the diffusion model can be considered an alternative way of storing a copy of those images” with the statutory definition of “Copies” as “being material objects… in which a work is fixed by any method now known or later developed, and from which the work can be perceived, reproduced, or otherwise communicated, either directly or with the aid of a machine or device.[iii]
Rather than make a determination as to whether the GenAI products contained infringing copies, the court struck the claims, but permitted them to be amended to provide further clarity which could, in turn, result in further pleading attacks.[iv]
In obiter dicta comments about the likely success of future challenges, the court suggested that the infringing status of the trained AI models (whether they contain copies or possibly derivative works of training images) might be inferred by whether the outputs reproduce only a few elements of the training images or whether synthetic images can be generated by text prompts that use an artist’s name.[v]
Whether GenAI outputs infringe copyrights
The defendants in the Anderson v Stability AI case also dispute the assertion that output images are infringing based on the facts pleaded in the Complaint. They rely heavily on the plaintiffs’ admission that, “[i]n general, none of the Stable Diffusion output images provided in response to a particular Text Prompt is likely to be a close match for any specific image in the training data.” They also attack the plaintiffs’ infringement theory that all outputs are necessarily infringing derivate works simply because they were derived from training images, some of which may be subject to copyright, and others which may not be subject to copyright. According, to this theory of the plaintiffs:
“Every output image from the system is derived exclusively from the latent images, which are copies of copyrighted images. For these reasons, every hybrid image is necessarily a derivative work.”
The court rejected the broad infringement theory that all outputs of a GenAI system must be infringing since some of the training images may not have been subject to copyright protection or be based on the copyright protected images:
A problem for plaintiffs… is that it is simply not plausible that every Training Image used to train Stable Diffusion was copyrighted (as opposed to copyrightable), or that all DeviantArt users’ Output Images rely upon (theoretically) copyrighted Training Images, and therefore all Output images are derivative images.
The court also rejected the broad infringement theory that all outputs of a GenAI system must be infringing on the basis that not every GenAI generated synthetic output would meet the US copyright requirements for derivative works. Under U.S. law, a derivative work must bear some similarity to the original work or contain the protected elements of the original work.[vi]
The defendants also sought to strike out the Plaintiffs’ claim that at least some training images were infringing copies or derivative works of training images. They did so, in part, based on the alleged admissions in the Complaint that “none of the Stable Diffusion output images provided in response to a particular Text Prompt is likely to be a close match for any specific image in the training data” and thus could not be substantially similar to any training images. The court commented on this stating:
This stands to reason: the use of conditioning data to interpolate multiple latent images means that the resulting hybrid image will not look exactly like any of the Training Images that have been copied into these latent images.
However, the court refused the strike these claims without leave to amend based on other parts of the Complaint that did allege substantial similarities between training images and synthetically derived images.[vii] This suggests the court accepted that such a well pleaded claim could be tenable.
Direct Infringement Contentions Against Midjourney
The plaintiffs’ claims with respect to Midjourney are similar to those against DeviantArt.[viii] Accordingly, the court made the same order it made with respect to these claims as it made with respect to DevianArt.
The claim against Midjourney also included the allegation that Midjourney was trained using a subset of the training images used to train Stable Diffusion. The court concluded there was insufficient facts regarding what training, if any, Midjourney conducted for its product. It therefore ruled that the plaintiffs needed to clarify their theory against Midjourney: “is it based on Midjourney’s use of Stable Diffusion, on Midjourney’s own independent use of training images to train the Midjourney product, or both?” The court permitted the Complaint to be amended to provide details of this training. It can thus be inferred that the court believed that using images for model training was a tenable pleading that could result in infringement (subject to applicable defenses).
As noted above, the summary judgment motions did not deal with the fair use defenses that the defendants intend to raise.
Kadry v Meta platforms
The decision in the Kadry v Meta is considerably shorter than the Stability AI decision. As in the Stability AI decision, the court dismissed all but one of the copyright claims with leave to amend the Copmplaint.
In Kadry, Meta also did not move to dismiss the claim that alleged that the unauthorized copying of the plaintiffs’ books for purposes of training its LLaMA GenAI system constitutes copyright infringement.
Whether the LLaMA Models are Infringing Derivative Works
According to the decision, the plaintiffs claim that the “LLaMA language models are themselves infringing derivative works” because the “models cannot function without the expressive information extracted” from the plaintiffs’ books. The court rejected this theory calling it “nonsensical”.
This is nonsensical. A derivative work is “a work based upon one or more preexisting works” in any “form in which a work may be recast, transformed, or adapted.” 17 U.S.C. § 101. There is no way to understand the LLaMA models themselves as a recasting or adaptation of any of the plaintiffs’ books.
The short decision did not canvass whether the LLAMA models are capable of infringing because they contain copies of the training images fixed within the trained models within the meaning of “Copies” in the U.S. Copyright Act.
Whether AI Outputs are Infringing Derivative Works
According to the court, the plaintiffs also claim that “every output of the LLaMA language models is an infringing derivative work,” and that because third-party users initiate queries of LLaMA, “every output from the LLaMA language models constitutes an act of vicarious copyright infringement.”
The court stated that “the complaint offers no allegation of the contents of any output, let alone of one that could be understood as recasting, transforming, or adapting the plaintiffs’ books.” Further, [W]ithout any plausible allegation of an infringing output, there can be no vicarious infringement.”
The court stated that it did not follow that because the books to which the plaintiffs owned the copyrights may have been copied that they did not need to establish any similarity between the books used for training and the LLaMA outputs.
The plaintiffs are wrong to say that, because their books were duplicated in full as part of the LLaMA training process, they do not need to allege any similarity between LLaMA outputs and their books to maintain a claim based on derivative infringement. To prevail on a theory that LLaMA’s outputs constitute derivative infringement, the plaintiffs would indeed need to allege and ultimately prove that the outputs “incorporate in some form a portion of” the plaintiffs’ books. Litchfield v. Spielberg, 736 F.2d 1352, 1357 (9th Cir. 1984); see also Andersen v. Stability AI Ltd., No. 23-CV-00201-WHO, 2023 WL 7132064, at *7-8 (N.D. Cal. Oct. 30, 2023) (“[T]he alleged infringer’s derivative work must still bear some similarity to the original work or contain the protected elements of the original work.”); 2 Melville B. Nimmer & David Nimmer, Nimmer on Copyright § 8.09 (Matthew Bender Rev. Ed. 2023) (“Unless enough of the pre-existing work is contained in the later work to constitute the latter an infringement of the former, the latter, by definition, is not a derivative work.”); 1 Melville B. Nimmer & David Nimmer, Nimmer on Copyright § 3.01 (Matthew Bender Rev. Ed. 2023) (“A work is not derivative unless it has substantially copied from a prior work.”)…
The court left open a claim by the plaintiffs that individual LLaMA outputs could be infringing derivative works. However, this claim would need to be pleaded as part of any amended Complaint.[ix]
Thomson Reuters v Ross Intelligence
The TR v Ross case, is the first U.S. decision to provide any guidance on whether the fair use defense might apply to otherwise infringing acts arising from the training, development and offering of GenAI products based on the unauthorized use of copyright content for GenAI training.
Background
The TR v Ross case is a much different case than many of the other cases that involve infringement of copyright claims related to the copying of books or other text based materials by GenAI systems. In the main, in these cases the defendants such Open AI or Meta are accused of infringement, among other things, by training their models using common crawl and massive datasets of infringing copies of books and other text based works taken from pirated datasets such as Books1, Books2, or Books3 (with claims that datasets are derived from Bibliotik which includes shadow libraries from the notorious sites LibGen, Z-Library and Sci-Hub) or from unauthorized archives of print publications such as the claims made by the New York Times.
In this case, the defendant Ross allegedly copied headnotes from TR’s Westlaw platform and, using a contractor, copied them to create memos with legal questions and answers which it used to train a GenAI system. Ross claimed that its “Bulk Memo Project” was designed to “avoid human intermediated materials.” Users would enter questions and its search engine would spit out quotations from judicial opinions and not copies of the Westlaw headnotes. TR argued, to the contrary, that Ross’ GenAI system was being designed to output materials copied from or derived from the Westlaw headnotes.
There were numerous legal issues canvassed in the decision. The most relevant for shedding some light on the liability associated with using copyright content for training a GenAI system is the court’s analysis of the U.S. fair use defense. The court refused to rule on the question because of numerous factual disputes. Despite this, the decision is of interest because of the analytical approach used by the court.
Fair Use
The parties cross-moved on Ross’s fair-use defense. The court explained that fair uses balances four factors:
(1) the purpose and character of the use, (2) the nature of the copyrighted work, (3) the amount and substantiality of the portion used in relation to the copyrighted work as a whole, and (4) the effect of the use upon the potential market for the copyrighted work. 17 U.S.C. § 107.
The first and fourth factors are most important. I therefore only summarize the court’s discussion of those two factors.
The Purpose and Character of the Use
The court explained that the first factor has two subparts: commerciality and transformativeness. These considerations interact:
The more transformative the new work, the less will be the significance of commercialism. Commerciality is straightforward: it asks whether the use was for profit. Transformativeness is less so: “a transformative use is one that communicates something new and different from the original or expands its utility, thus serving copyright’s overall objective of contributing to public knowledge.
The court found that Ross’s “uses were undoubtedly commercial” as one of its goals was to compete with Westlaw. TR contended that this commercial use weighed heavily against finding fair use. In support of this, it relied on the recent the Supreme Court decision in Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 143 S. Ct. 1258 (2023) where the U.S. Supreme Court determined that the use in question was not fair largely by emphasizing its commercial nature.
The court declined, however, to directly apply that on the basis that “use’s transformativeness may outweigh its commercial character” and that in Warhol, “both elements point[ed] in the same direction.” The court referred to the recent US Supreme Court decision in Google LLC v. Oracle Am., Inc., 141 S. Ct. 1183 (2021) where the court “placed much more weight on transformation than commercialism”. The court then focused on transformativeness summarizing the arguments of the parties as follows:
Thomson Reuters paints a black-and-white picture on transformativeness: Westlaw is a legal-research platform that synthesizes the law; Ross used Westlaw’s syntheses to build a legal-research platform that also synthesizes the law. Ross, on the other hand, presents a more nuanced account: Westlaw headnotes and key numbers annotate opinions for users. Ross wanted to build a search engine that “avoids human intermediated materials,” meaning a user would simply enter a query and get a responsive quotation from a judicial opinion, no clicking around or commentary needed. Though Ross and Westlaw both help answer legal questions, Ross says it transformed the Westlaw headnotes beyond recognition.
Ross describes its process of transforming the Bulk Memos like this: First, it receives the Bulk Memos in its database. Then, it converts the plain-language entries into numerical data. Next, it feeds that data into its machine-learning algorithm to teach the artificial intelligence about legal language. The idea is that the artificial intelligence will be able to recognize patterns in the question-answer pairs. It can then use those patterns to find answers not just to the exact questions fed into it, but to all sorts of legal questions users might ask…
Ross says its AI studied the headnotes and opinion quotes only to analyze language patterns, not to replicate Westlaw’s expression. So the translation was only a minor step in a broader, transformative use…
TR argued that Ross’ actions were rather focused on the use of headnotes to get its GenAI product to replicate Westlaw’s headnotes.
The court concluded there were material questions of fact that needed to go to a jury. On the first factor, the court opined that if Ross’s AI only studied the language patterns in the headnotes to learn how to produce judicial opinion quotes that would tend to favour fair use. On the other hand if Ross used the untransformed text of headnotes to get its GenAI product to replicate and reproduce the creative drafting done by Westlaw’s attorney-editors, that would tend to lead to a contrary finding.[x]
The Effect of the Use Upon the Market for the Work
The Fourth factor in the fair use framework asks whether the use has a “meaningful or significant effect” on the value of the original or its potential market. “This inquiry must take account not only of harm to the original but also of harm to the market for derivative works.”
According to the court, since copyright only protects expression in a work, if “the source of the loss is not that the original’s expression is being appropriated, “the type of loss of sale envisioned above will generally occur in relation to interests that are not protected by the copyright.” The court also stated that “transformativeness feeds into this factor as well. “[T]he more the copying is done to achieve a purpose that differs from the purpose of the original, the less likely it is that the copy will serve as a satisfactory substitute for the original.” “Finally, according to the court, in evaluating market impact, courts must pay special attention to “the realities of how technological works are created and disseminated.”
Here again the court determined it could not resolve this issue based on disputed facts.
Here, those “realities” are disputed. Thomson Reuters claims three potential markets, but they boil down to two: the market for Westlaw itself as a legal research platform and the market for its data. It says Ross’s plan all along was to create a substitute for Westlaw. And it says that this plan worked, as some Ross customers cancelled their Westlaw subscriptions. As for the market for its data, Thomson Reuters says there is a traditional licensing market and a burgeoning one for AI training data. It argues that it lost traditional licensing revenue because Ross obtained Westlaw content through LegalEase. And it suggests that there is a potential market for Westlaw’s training data; after all, Ross paid LegalEase over a million dollars for the Bulk Memos. That burgeoning market would be harmed by copying like Ross’s.
One fact is undisputed here: Ross and Thomson Reuters both compete in the market for legal research platforms. But that alone does not reveal whether Ross’s AI product is a substitute for Westlaw. Ross’s use might be transformative, creating a brand-new research platform that serves a different purpose than Westlaw. If so, it is not a market substitute. Ross also argues that Thomson Reuters has never participated—and would never participate—in this market for its training data. Because a reasonable jury could find for either side on these factual market-impact questions, I cannot resolve them at summary judgment.
The Court made a similar finding as to how to weigh the public benefits associated with using copyright materials without a license to train a new AI system.
Finally, “we must take into account the public benefits the copying will likely produce.” Google, 141 S. Ct. at 1206. And “we are free to consider the public benefit resulting from a particular use notwithstanding the fact that the alleged infringer may gain commercially.” Sega, 977 F.2d at 1523. This “[p]ublic benefit need not be direct or tangible, but may arise because the challenged use serves a public interest.” Id.
The parties provide competing narratives of public benefit. Ross’s research platform might increase access to the law at a lower cost. Or it might just reduce the incentives for Thomson Reuters, and similarly situated entities, to create content like headnotes in the future.
Deciding whether the public’s interest is better served by protecting a creator or a copier is perilous, and an uncomfortable position for a court. Copyright tries to encourage creative expression by protecting both. Here, we run into a hotly debated question: Is it in the public benefit to allow AI to be trained with copyrighted material?
The value of any given AI is likely to be reflected in the traditional factors: How transformative is it? Can the public use it for free? Does it discourage other creators by swallowing up their markets? So an independent evaluation of the benefits of AI is unlikely to be useful yet, even though both the potential benefits and risks are huge. Suffice it to say, each side presents a plausible and powerful account of the public benefit that would result from ruling for it. So a jury must decide the fourth factor—and the ultimate conclusion on fair use.
Getty Images v Stability AI
This UK proceeding includes a claim for copyright infringement by Getty Images against Stability AI’s deep learning GenAI model Stable Diffusion and its commercial product Dream Studio. The gravamen of the complaints in the UK proceeding are:
- That the use of Getty licensed content including photographs, video footage and illustrations that are protected as original artistic or film works under the UK Copyright, designs, and Patents Act (the CDPA) that were scraped from Getty websites and used to train Stable Diffusion infringed copyright.
- That the output of Stable Diffusion in the form of synthetic images (accessed by users in the UK) is infringing in that they reproduce a substantial part of the Getty copyright works and the dissemination of the output to UK users constitutes a communication to the public under UK copyright law.
- That the trained Stable Diffusion software infringed copyright and that its distribution, importation, offering for sale and other dealings in the UK constituted secondary infringement under the CDPA.
- That the image-to-image feature of Stable Diffusion which allegedly permits users to make essentially identical copies of the Getty copyright works infringes copyright.
There were several interlocutory motions brought by the parties which were the subject of the decision. However, Stability AI did not contest in these motions the claim of copyright infringement by reason of the copying and communication to the public where Stable Diffusion is used to generate an image that reproduces the whole or a substantial part of the Getty copyright works contrary to the CDPA. This claim relies upon access to Stable Diffusion by users in the UK. Unless settled, this claim will proceed to trial in the UK.
Whether use of the Getty Images to Train or Develop Stable Diffusion is Infringing.
Stability AI did not challenge in the motions that the copying of the Getty Images to train Stable Diffusion could implicate the reproduction right. The challenge in the motion was whether, as a factual matter, the training or development, or at least some of it, took place in the UK. The location of any alleged copying is important because, and it was common ground on the motion, that copyright is a territorial right which confers protection on its holder only within the territory of the allegedly infringing act, in this case, the United Kingdom.[xi]
The court reviewed the substantial evidence adduced on the summary judgment motion as to where the development and training of Stable Diffusion took place and was unable to conclude that the training and development claim had no real prospect of success. It therefor dismissed this motion for summary judgment.
Whether Making Dream Studio Available in the UK Constitutes Secondary Infringement
Getty Images claimed that the website through which Stability AI sells access to Dream Studio to UK users (which enables users to generate images) and the fact that Stable Diffusion is available to download on an open-source basis to third parties in the UK via GitHub constitutes secondary infringement under the CDPA.
These claims were premised on what appeared to be an uncontested issue on the motion, namely that the trained Dream Studio model could contain infringing reproductions of the Getty copyright works in a material form.
What was in issue on the motion was whether Dream Studio could be considered an “article” for the purposes of secondary infringement under the CDPA.[xii] Under UK law, secondary infringement can be made out, among other things, if in certain circumstances, a person who, without the licence of the copyright owner sells or offers or exposes for sale, or in the course of a business exhibits in public or distributes, or distributes to such an extent as to affect prejudicially the owner of the copyright, an article which is, and which he knows or has reason to believe is, an infringing copy of the work. An article may also be an infringing copy if it has been or is proposed to be imported into the UK, and its making in the UK would have constituted an infringement of the copyright in the work in question, or a breach of an exclusive licence agreement relating to that work.[xiii]
Stability AI argued that Dream Studio was intangible property and could not be an “article” for the purposes of UK secondary infringement principles.[xiv] The court was not satisfied that this issue was clear enough to be settled on the summary judgment motion and postponed this issue to a trial.[xv]
Whether an Image-to-Image Feature can be Infringing
Getty Images brought a motion to amend its application to allege that an image-to-image feature which enables Stable Diffusion to generate a synthetic image output in response to an image uploaded by a user (either with or without a text prompt) was also infringing. This feature enables a user to determine how closely the synthetic image output matches the image prompt by the use of an “image strength” slider, with the maximum image strength and/or values approaching the maximum image strength providing images which comprise the whole, or a substantial part, of the image prompt.[xvi]
Stability AI opposed the amendment motion arguing, among other things, that it had no real prospect of success at trial. Stability AI’s main argument appeared to be that when a copy of Stable Diffusion is downloaded by a user from GitHub with a view to then entering the image prompt, it is in fact the user who brings about the copying (i.e. by voluntarily entering into Stable Diffusion a copyright work that the user has downloaded from the Getty Images websites) and not the defendant.
The court ruled that based on the arguments as they were presented in court, the image-to-image claim as pleaded had a real prospect of success and that, having regard to the balancing exercise which is required in pleading motions, it was a claim that should go to trial. The judge’s reasons, in summary, were that it was unclear that Stability AI’s role was merely passive and instrumental (the attributes of a neutral intermediary) and that since the claim was similar to the text-to-image claim, both claims should proceed to trial.[xvii]
Summary and Conclusions on GenAI and Copyright Infringement
At the beginning of this post, I noted that the decisions rendered to date had canvassed several very important questions about whether the creation and use of GenAI systems infringe copyright. Below, I set out a summary of the answers provided by the courts to date. However, before doing so I repeat my caution above, that all decisions to date should be considered as preliminary, having been rendered on summary judgment or pleading motions and without full argument on the facts and law after a trial. Further, the decisions are only based on the allegations made by the plaintiffs in the cases before the courts, the defenses raised by the defendants in these cases, and arise from the specific GenAI systems in issue. Thus, any conclusions or inferences that can be drawn from these decisions must therefore be understood as subject to these caveats.
However, subject to the above caveats, here is a summary of what the courts have suggested may be tenable pleadings which could, depending on the facts, result in copyright infringement or defenses to infringement. Note my use of the phrase “tenable pleading” does not suggest this pleading would be a successful one. Rather it is a properly pleaded claim that may be sufficient to create a triable issue for trial.
Whether downloading and copying of works (or other copyright subject matter) to train GenAI models infringes copyright – No defendant in the cases canvassed has appeared to claim that the training activity did not engage the exclusive right to reproduce or copy a work. (However, the defendants claim such copying is subject to one or more defenses such as fair use.)
Whether the trained GenAI models are infringing because they contain or embody copies of training materials in a material object or material form or are, under US law, derivative works – The decisions appear to accept or implicitly accept as a tenable pleading that GenAI models may contain infringing copies (“compressed copies”) of copyright protected training materials. However, the courts have not explained the legal basis for their findings or connected their reasons to the U.S. definition of “Copies”. The U.S. courts do not accept as a tenable pleading that GenAI systems are derivative works of copyrighted training materials simply because they are derived from these training materials.
Whether all outputs from GenAI systems are derivative works because they are derived from copyright works and whether specific outputs can infringe the copyrights in specific training materials – The U.S. courts have so far rejected as a tenable pleading the theory that all GenAI outputs must be derivative works because they are “derived” from copyright protected training materials (as well as materials not subject to copyright). However, the courts appear to accept that an allegation that specific GenAI outputs reproduce specific training materials could be a tenable pleaded claim of infringement.
Whether outputs generated from an GenAI system located in the U.S. but accessed by users outside the US can infringe the making available right – This claim was not challenged in the UK Stability Getty Images case. Accordingly, it appears to represent a tenable pleading.
Whether offering a GenAI system for download from the U.S. to users outside the U.S. such as in the UK could constitute secondary infringement under the law of the country of access – This claim is being advanced in the Stability AI Getty Images case after the UK court declined to strike it out. But, note the secondary infringement claim did include an allegation that providing access to Stable Diffusion to users in the UK constitutes secondary infringement.
Whether fair use is potentially available in the U.S. as a defense to GenAI copyright infringement claims – Unsurprisingly, whether fair use as a defense to copyright infringement applies is a very factual determination based on an analysis that includes the four fair use factors under the U.S. Copyright Act and public policy considerations. The TR v Ross case emphasized the importance of considering whether the GenAI system is transformative under the first factor and the market-impact considerations under the fourth factor including the “transformativeness” of the GenAI product and “the realities of how technological works are created and disseminated”.
I hope you have found this summary of the recent decisions useful and informative. I am sure there will be more to come in 2024.
For other posts on GenAI and copyright, you may be interested in:
- Do generative AI inventions and works qualify for patents and copyrights? The Thaler and SURYAST decisions
- AI, Copyright + the Wisdom of Experience: my keynote speech at All In
- OpenAI comes out swinging in motions to dismiss copyright class action claims
- Generative AI and copyrights: My talk to the IPO
- Copyright does not protect content produced by Generative AI (GenAI): Thaler v Perlmutter
[i] Direct Copyright Infringement, 17 U.S.C. § 106; (2) Vicarious Copyright Infringement, 17 U.S.C. § 106; (3) violation of the Digital Millennium Copyright Act, 17 U.S.C. §§ 1201-1205 (“DMCA”); (4) violation of the Right to Publicity, Cal. Civil Code § 3344; (5) violation of the Common Law Right of Publicity; (6) Unfair Competition, Cal. Bus. & Prof. Code § 17200; and (7) Declaratory Relief. Plaintiffs also assert a breach of contract claim against DeviantArt only.
[ii] “Plaintiffs have adequately alleged direct infringement based on the allegations that Stability “downloaded or otherwise acquired copies of billions of copyrighted images without permission to create Stable Diffusion,” and used those images (called “Training Images”) to train Stable Diffusion and caused those “images to be stored at and incorporated into Stable Diffusion as compressed copies.”
[iii] 17 U.S.C. &s;§ 101 defined “Copies” as “material objects, other than phonorecords, in which a work is fixed by any method now known or later developed, and from which the work can be perceived, reproduced, or otherwise communicated, either directly or with the aid of a machine or device. The term “copies” includes the material object, other than a phonorecord, in which the work is first fixed.” See, MAI Systems Corp. v. Peak Computer Inc., 26 U.S.P.Q. 2d 1458 (9th Cir. 1993).
[iv] “Plaintiffs will be required to amend to clarify their theory with respect to compressed copies of Training Images and to state facts in support of how Stable Diffusion — a program that is open source, at least in part— operates with respect to the Training Images. If plaintiffs contend Stable Diffusion contains “compressed copies” of the Training Images, they need to define “compressed copies” and explain plausible facts in support. And if plaintiffs’ compressed copies theory is based on a contention that Stable Diffusion contains mathematical or statistical methods that can be carried out through algorithms or instructions in order to reconstruct the Training Images in whole or in part to create the new Output Images, they need to clarify that and provide plausible facts in support.”
[v] “Depending on the facts alleged on amendment, DeviantArt (and Midjourney) may make a more targeted attack on the direct infringement contentions. It is unclear, for example, if Stable Diffusion contains only algorithms and instructions that can be applied to the creation of images that include only a few elements of a copyrighted Training Image, whether DeviantArt or Midjourney can be liable for direct infringement by offering their clients use of the Stable Diffusion “library” through their own apps and websites. But if plaintiffs can plausibly plead that defendants’ AI products allow users to create new works by expressly referencing Anderson’s works by name, the inferences about how and how much of Anderson’s protected content remains in Stable Diffusion or is used by the AI end-products might be stronger.”
[vi] The court stated the following in regards to the allegation that every output derived in some way from a training image must be infringing:
Even if that clarity is provided and even if plaintiffs narrow their allegations to limit them to Output Images that draw upon Training Images based upon copyrighted images, I am not convinced that copyright claims based a derivative theory can survive absent “substantial similarity” type allegations. The cases plaintiffs rely on appear to recognize that the alleged infringer’s derivative work must still bear some similarity to the original work or contain the protected elements of the original work. See, e.g., Jarvis v. K2 Inc., 486 F.3d 526, 532 (9th Cir. 2007) (finding works were derivative where plaintiff “delivered the images to K2 in one form, and they were subsequently used in the collage ads in a quite different (though still recognizable) form. The ads did not simply compile or collect Jarvis’ images but rather altered them in various ways and fused them with other images and artistic elements into new works that were based on— i.e., derivative of—Jarvis’ original images.”) (emphasis added); ITC Textile Ltd. v. Wal-Mart Stores Inc., No. CV122650JFWAJWX, 2015 WL 12712311, at *5 (C.D. Cal. Dec. 16, 2015) (“Accordingly, even if Defendants did modify them slightly, such modifications are not sufficient to avoid infringement in a direct copying case. . . . Thus, the law is clear that in cases of direct copying, the fact that the final result of defendant’s work differs from plaintiff’s work is not exonerating.”) (emphasis added); see also Litchfield v. Spielberg, 736 F.2d 1352, 1357 (9th Cir. 1984) (“a work is not derivative unless it has been substantially copied from the prior work”); Authors Guild v. Google, Inc., 804 F.3d 202, 225 (2d Cir. 2015) (“derivative works over which the author of the original enjoys exclusive rights ordinarily are those that re-present the protected aspects of the original work, i.e., its expressive content”).
[vii] “But other parts of plaintiffs’ Complaint allege that Output Images can be so similar to plaintiff’s styles or artistic identities to be misconstrued as “fakes.” Compl. ¶¶ 44ii, 171-172. Once plaintiffs amend, hopefully providing clarified theories and plausible facts, this argument may be re-raised on a subsequent motion to dismiss.”
[viii] The claims with respect to Mijourney are “(1) direct infringement by reproducing protected works by scraping (i.e., copying) and reproducing images used as training images for the Midjourney Product; (2) direct infringement by distributing Stable Diffusion, which contains compressed copies of the training images, as part of the Midjourney Product; (3) direct infringement by creating and distributing their Midjourney Product, which is itself an infringing derivative work; and (4) generating and distributing output images which are infringing derivative works.”
[ix] “That’s the same thing the plaintiffs would need to do here with respect to the content of LLaMA’s outputs. To the extent that they are not contending LLaMa spits out actual copies of their protected works, they would need to prove that the outputs (or portions of the outputs) are similar enough to the plaintiffs’ books to be infringing derivative works. And because the plaintiffs would ultimately need to prove this, they must adequately allege it at the pleading stage.”
[x] “So whether the intermediate copying caselaw tells us that Ross’s use was transformative depends on the precise nature of Ross’s actions. It was transformative intermediate copying if Ross’s AI only studied the language patterns in the headnotes to learn how to produce judicial opinion quotes. But if Thomson Reuters is right that Ross used the untransformed text of headnotes to get its AI to replicate and reproduce the creative drafting done by Westlaw’s attorney-editors, then Ross’s comparisons to cases like Sega and Sony are not apt. Again, this is a material question of fact that the jury needs to decide.”
[xi] According to the court referring to the copyright territoriality principle:“”As Lord Neuberger PSC said in Starbucks (HK) Ltd v British Sky Broadcasting Group plc [2015] UKSC 31 at [59], approving Professor Wadlow in The Law of Passing-off – Unfair Competition by Misrepresentation 4th ed, 2011, para 3-131: ” the nature of copyright as a legal right of property arising in any given jurisdiction from national legislation, common law or self-executing treaty, means that it must be wrong to speak as if there were a single international copyright”.”
[xii] The secondary infringement claims were summarized as follows by the court:
Paragraphs 52.4 and 52.5 are in the following terms:
“52.4 As regards the acts identified above at paragraphs 44 to 46, importing into the United Kingdom, otherwise than for private and domestic use, an article, namely Stable Diffusion, which is and which the Defendant knows or has reason to believe is, an infringing copy of the Copyright Works or each or any of them, contrary to section 22 of the CDPA; and/or
52.5 As regards the acts identified above at paragraphs 44 to 46, possessing in the course of a business, selling or letting for hire, or offering or exposing for sale or hire, an article, namely Stable Diffusion, which is and which the Defendant knows or has reason to believe is, an infringing copy of the Copyright Works or each or any of them, contrary to section 23 of the CDPA”.
Paragraph 53 pleads that “pending disclosure and/or evidence” the Claimants’ case that Stable Diffusion is an infringing copy of the Copyright Works is dependent either upon the fact (paragraph 53.1) that the making of Stable Diffusion constituted an infringement (i.e. the Development and Training Claim) or the fact that (paragraph 53.2) Stable Diffusion “has been imported into the United Kingdom” and by reason of the manner in which it was trained as set forth in paragraph 47 of the Particulars of Claim “its making in the United Kingdom would have constituted an infringement” of the copyright in the Copyright Works.
Paragraph 54 pleads particulars of knowledge, said by the Defendant in its skeleton argument to be insufficiently particularised, but not pursued at the hearing.
[xiii] The key secondary infringement provisions of the CDPA are in the following terms:
“22. Secondary infringement: importing infringing copy
The copyright in a work is infringed by a person who, without licence of the copyright owner, imports into the United Kingdom, otherwise than for his private and domestic use, an article which is, and which he knows or has reason to believe is, an infringing copy of the work.
23.. Secondary infringement: possessing or dealing with infringing copy
The copyright in a work is infringed by a person who, without the licence of the copyright owner-
(a) Possesses in the course of a business,
(b) Sells or lets for hire or offers or exposes for sale or hire,
(c) In the course of a business exhibits in public or distributes, or
(d) Distributes otherwise than in the course of a business to such an extent as to affect prejudicially the owner of the copyright, an article which is, and which he knows or has reason to believe is, an infringing copy of the work.
27.. Meaning of “infringing copy”
(1) In this Part “infringing copy” in relation to a copyright work, shall be construed in accordance with this section.
(2) An article is an infringing copy if its making constituted an infringement of the copyright in the work in question
(3) An article is also an infringing copy if-
(a) it has been or is proposed to be imported into the United Kingdom, and
(b) its making in the United Kingdom would have constituted an infringement of the copyright in the work in question, or a breach of an exclusive licence agreement relating to that work.
[xiv][xiv] Stability AI’s argument that secondary infringement requires that an article be a tangible article was summarixed by the court as follows:
There is no relevant statutory definition of “article”. However, the Defendant’s case is that, on a true interpretation of the statutory provisions to which I have referred, an article can only be something which is capable of being “imported” (section 22); “possessed” (section 23(a)); “distributed” (section 23(c)-(d)) and “made” (section 27), i.e. a physical tangible thing. The Defendant’s skeleton notes that the only cases that counsel has been able to identify under sections 22 and 23 all concern physical tangible articles, such as pairs of shoes (LA Gear Inc v Hi-Tec Sports plc [1992] FSR 121); compact discs (Springsteen v Flute International Ltd [1999] EMLR 180) and football sticker albums (FAPL v Panini UK Ltd [2004] 1 WLR 1147).
The Defendant points out that it is settled law that abstract information is not property and cannot be the subject of a claim for detinue or conversion, relying in particular upon the following observations made by Floyd LJ in Your Response Limited v Datateam Business Media Limited [2014] EWCA Civ 281 at [42]:
“An electronic database consists of structured information. Although information may give rise to intellectual property rights, such as database right and copyright, the law has been reluctant to treat information itself as property. When information is created and recorded there are sharp distinctions between the information itself, the physical medium on which the information is recorded and the rights to which the information gives rise. Whilst the physical medium and the rights are treated as property, the information itself has never been.”
and by Arnold LJ in Thaler v Comptroller General of Patents, Trade Marks and Designs [2021] EWCA Civ 1374 at [125]:
“There are good reasons of principle and policy for this rule: it would have alarming consequences if there was property in information”.
Thus, submits the Defendant, whilst cases relating to whether there is property in information do not bind this court when it comes to the interpretation of sections 22 and 23 of the CDPA, those cases suggest that it is necessary to look to dealings in tangible property (goods or merchandise) which embody information when assessing the “article” for the purposes of secondary infringement. The Defendant contends that examples of this approach may be seen in Sony Computer Entertainment Inc v Ball [2005] FSR 9 and in Wheat v Google LLC [2018] EWHC 550 (Ch) to which I shall return in a moment.
Finally, the Defendant contends that this question of statutory construction is a straightforward question of law, that the Claimants have no real prospect of establishing that “article” includes intangible articles and that the point should be determined (whether in its favour or not) at this hearing.
[xv] The court’s reasons for considering that an “article” might include an intangible item included the following:
First, I am not satisfied on the arguments that I have heard that the Claimants are bound to fail on this point. I say that for the following reasons:
Sony v Ball
- Contrary to the Defendant’s submissions, I do not consider that Sony v Ballprovides the answer. In brief summary only, it was common ground in that case that a silicon chip was capable of being an article. However, a key question for the court was whether a chip that contained an infringing copy of digital data for only a small fraction of a second could be an article given its “ephemeral” state. It was argued that the chip could not be regarded as a “tangible substance” in such circumstances. Laddie LJ did not grapple with the question of tangibility. Rather, he held (at [15]) that there is nothing in the legislation to suggest that an object containing a copy of a copyright work, even if only ephemerally, is for that reason to be treated as not being an article. He went on to say “[o]n the contrary, the definition in s.27 points to the instant of making the copy as crucial to the determination of whether or not it is an infringing article. An article becomes an infringing article because of the manner in which it is made. Whether it is an infringing article within the meaning of the legislation must be determined by reference to that moment.”
- Where the judge simply did not address the argument as to whether the chip (or indeed the information it held) could be regarded as ‘tangible’, I do not agree with the Defendant that it is necessarily implicit in this reasoning that the information stored on the chip could not be an article, or indeed that an article is limited purely to “tangible” objects. The court only addressed the issue of whether the silicon chip was an article in circumstances where the life of the copy on the chip was of extremely limited duration. I reject the suggestion that it is possible to extrapolate from this decision the conclusion that an article must inevitably be a tangible thing.
Wheat v Google
- True it is that in Wheat v Google, the Chief Master expressed the view (at [48]) that “the acts referred to in section 23 CDPA are tangible acts of physical exhibition or distribution, not electronic transmission of information”, a view which appears supportive of the Defendant’s submission that an article can only be a physical tangible thing. However, as Ms Lane correctly points out, the hearing before the Chief Master concerned applications by the claimant (acting in person) to serve Google out of the jurisdiction. The Chief Master does not appear to have heard any detailed argument on this point – certainly he does not set any out – and he provides no reasoning in support of the view he has taken. He does not appear to have been invited to consider any authority or to engage in a textual or contextual analysis of the CDPA. Furthermore, he appears to have been concerned only with acts of “exhibition” and “distribution” and not with any of the additional acts of secondary infringement alleged in this case. In the circumstances, and with the greatest of respect to the Chief Master who appears to have received no assistance whatever on the point, I cannot attach any serious weight to his decision, which is not, in any event, binding on me.
The full statutory context
- I consider there to be two points about the statutory context which may potentially be relevant but in respect of which I have not heard full argument at the hearing and/or cannot determine that the Claimants are bound to fail.
- The firstconcerns the potential relationship between sections 22 and 23 and section 17 of the CDPA. It is clear from section 17(2) that reproduction of the copyright work (for the purposes of primary infringement) can be in any material form, including “storing the work in any medium by electronic means”. Ms Lane points out that there is no dispute for the purposes of the SJ Application that the making of Stable Diffusion involved reproducing the Copyright Works in a material form contrary to section 17 of the CDPA and she submits that this is of significance in the context of construing the meaning of “infringing copy” in section 27 and “article” in sections 22, 23 and 27. At first blush, I tend to agree, not least because in Sony v Ball, Laddie LJ appears to have taken the approach (at [16]-[17]) that when identifying an infringing copy for the purposes of section 27, it is necessary to have regard to what may be an infringing copy for the purposes of section 17. In particular, he observed that “[i]t would produce an unwarranted inconsistency in the [CDPA] were that material form [i.e., the chip containing the reproduction of the copyright work] not to be considered an article for the purpose of s.27”.
- Ms Lane referred me to a post Brexit EU authority to the effect that storage in a medium by electronic means can include cloud storage (see C-433/20, EU:C:2022:217Austro-Mechana Gesellschaft zur Wahrnehmung mechanisch-musikalischer Urheberrechte GmbH v Strato AG[2022] ECDR 10 at [15]-[18]), submitting (by reference to Warner Music UK Ltd v TuneIn Inc [2021] EWCA Civ 441 at [90]-[91] per Arnold LJ) that this judgment is highly persuasive, albeit not binding.
- Mr Saunders initially suggested in response to a question from me, that a “medium” (as identified in section 17) is similar to an “article”, albeit that he did not appear to maintain that point. He acknowledged that he had not focussed on section 17 in connection with the interpretation of sections 22 and 23 of the CDPA and indicated that he might wish to come back to the point. However, despite Ms Lane dealing with the point in some detail, Mr Saunders did not revisit it, and (although the Defendant sought to distinguish between primary infringement in section 17 and secondary infringement in sections 22 and 23 of the CDPA in its skeleton argument) I have no clear explanation from the Defendant as to why section 17 is not an important and potentially relevant part of the contextual analysis. Albeit in a slightly different context, Mr Saunders showed me a passage from the decision of the Competition Appeal Tribunal in Umbrella Interchange Fee Claimants v Umbrella Interchange Fee Defendants[2023] CAT 49, in which the Tribunal decided (for detailed reasons) not to follow existing EU law. Mr Saunders acknowledged that he could not possibly invite me at this hearing (in so far as it may be relevant) to determine whether this court should follow post Brexit EU authority.
- The secondarises by reason of a point made by Ms Lane in her oral submissions but which did not appear in the Claimants’ skeleton argument or its supplemental skeleton. Ms Lane took me in some detail to section 27(6) of the CDPA which provides that an “infringing copy” will include “a copy falling to be treated as an infringing copy” by virtue of a variety of other sections of the CDPA (introduced by way of amendment), there identified. She submits, correctly, that an essential part of the exercise of statutory construction is the consideration of the relevant words or phrases in their statutory context. She points to various provisions in these other sections which she submits (i) cast light on the meaning of “infringing copy” and, in particular, suggest that an “infringing copy” can include an “electronic” copy and, thus, an intangible copy; and (ii) suggest that not all of the secondary acts of infringement in section 23 are acts which must necessarily be associated with a tangible article – e.g. the act of selling or letting for hire.
- In his reply, Mr Saunders (who had not previously been alerted to this point and was obliged to deal with it “on his feet”) dealt with it only briefly. In essence he submitted that the Statutory Instruments that introduced the amendments to the primary legislation cannot have been intended to affect the true construction of existing provisions. However, he was unable to say whether there is any authority about the extent to which one can refer to later amendments to a statute to construe a pre-existing provision. This caused me at the time to observe that the matter did not appear to be straightforward and to express concern that the court did not have the material it needed to decide the issue. Mr Saunders responded that his primary position was that the argument had been raised too late and that he had not had an opportunity to look at it.
- Be that as it may, it appears to me that it would be unsafe finally to decide an important point of statutory construction in such circumstances. Whatever the rights and wrongs of the Claimants raising this argument at the eleventh hour, it has been raised and the Defendant was not in a position fully to deal with it at the hearing. Mr Saunders continued to suggest that the sections to which I was taken by Ms Lane cannot possibly be relevant to the interpretation of the words “infringing copies” or “article”, but I am not satisfied on the limited argument that I have heard that the Claimants have no real prospect of arguing successfully to the contrary at trial.
[xvi] The specific claims of infringement related to the image-to-image feature was described a follows:
The relevant paragraphs are 50A-50B.3 of the draft Amended Particulars of Claim:
“50A. From a date unknown but since at least in or around March 2023 the Defendant introduced an image-to-image feature which enables Stable Diffusion to generate a synthetic image output in response to an image uploaded by a user (either via Dream Studio or via a copy of Stable Diffusion downloaded from GitHub or via commercial offerings using Stable Diffusion), either with or without a text command. The user is able to determine how closely the synthetic image output matches the image prompt by the use of an “image strength” slider, with the maximum image strength and/or values approaching the maximum image strength providing images which comprise the whole or a substantial part of the image prompt.
50B. Accordingly, when a user provides an image prompt (either via Dream Studio or via a copy of Stable Diffusion downloaded from GitHub or via commercial offerings using Stable Diffusion) comprising one of the Copyright Works or the Visual Assets and uses the maximum image strength or a value approaching the maximum image strength (the precise value to be identified by way of disclosure and/or evidence), the synthetic image output comprises the whole or a substantial part of the Copyright Work or Visual Asset. Pending the provision of disclosure and/or evidence, the Claimants rely upon the following facts and matters:
50B.1. In the case of Stable Diffusion downloaded from GitHub, the image strength is a value between 0.0 and 1.0 that controls the amount of noise that is added to the image prompt. As explained in the guidance on the GitHub website (a copy of which is attached hereto at Annex 8A) values that approach 1.0 allow for lots of variations in the synthetic image output by removing all pixel-based information. By contrast, values that approach 0.0 (i.e. the maximum image strength) result in a synthetic image output that reproduces the whole or a substantial part of the image prompt.
50B.2. When an image is uploaded by a user, it is first converted into noise, and this noise pattern is used as an input to the image synthesis process. Thus, the image-to-image feature uses both Stable Diffusion and the pixels of the user-supplied, uploaded image.
50B.3. Copies of synthetic images that were generated in response to image prompts comprising a Copyright Work or Visual Asset, together with the image strength indicated, uploaded via a copy of Stable Diffusion downloaded from GitHub are attached hereto at Annex 8B”.
- Cross-references to these paragraphs then appear in amended versions of paragraphs 52.2, 52.3 and 55.2 making express the plea of copyright infringement (by communication to the public or by authorising acts of reproduction or communication to the public, contrary to sections 16(2) and 20 CDPA) and database right infringement (by re-utilising a substantial part of the contents of the Database). Additional amendments to paragraphs 33 and 39 plead that the use of Stable Diffusion to generate synthetic images in response to image prompts commenced on “a date unknown but since at least in or around March 2023”.
[xvii] The court’s reasons for permitting the image-to-image claim amendment was expressed as follows:
The claim of copyright infringement as advanced is based on more than the Defendant merely acting as intermediary, or host (cf. Google France SARL v Louis Vuitton Malletier SA C-236/08 EU:C:2010:159 at [55]-[58] in the context of trade mark infringement). As is pleaded in 50B.2, and as is confirmed by Professor Farid, an image which is entered into the Stable Diffusion software is ingested into the model, is pixelated, made into a noisy image and is then used to create part of the output. In other words, the software is reacting to, and engaging with, the image that is entered by the user, thereby undertaking a process which leads to the generation by the software of output in the form of an image. This factual case, which is said to apply regardless of whether the image strength is set to a value that allows for lots of variations in the synthetic image output or whether it is set to a value approaching the maximum image strength, is not challenged by the Defendant.
Upon close analysis of the technology at trial, the factual position may prove to be different. It may be possible to establish that, as Mr Saunders submits, the user who inputs the image (particularly where the image strength slider is set to produce a maximum image strength) is “doing the interaction” and that Stable Diffusion is really playing an immaterial part. However, I cannot possibly determine that factual technical issue at present and I note that in its skeleton argument the Defendant was unable to go further than to say that its software “has little to do” with the creation of the output image, a tacit admission that it at least has something to do with it. Resolution of this issue will require the court to hear detailed evidence as to the workings of the Stable Diffusion software, including the extent to which (if at all) the positioning of the image strength slider affects the way in which the software functions. For the purposes of the amendment application, I certainly cannot determine at this stage that the claim as pleaded has no real prospect of success.
Furthermore, the points I have made at 109 and 110 above appear to me to place this claim (at least on the facts as presently pleaded) in very similar territory to the text-to-image claim. Where there is no challenge to the text-to-image claim and it is accepted by the Defendant that that claim must go to trial, I agree with Ms Lane that it would be rather odd if the court were to find itself in a position where it could not also determine the image-to-image claim, which relies upon the same technology and a very similar mechanism for generating images – the only difference being the nature of the prompt required for the generation of the image. I disagree with Mr Saunders that the proposed amendments necessarily concern a “fundamentally different kind of case”, or, at least, that that is a conclusion that the court is able to reach at this stage.
As for the question of whether the “local” nature of the activity (when there is a downloading of Stable Diffusion from GitHub onto a user’s personal computer) negates any prospect of the Claimants establishing a case of infringement, the Claimants say that they have a sufficiently arguable case of (i) communication to the public (which they say includes communication to a “new public” in the form of users; i.e. not the public that the Claimants originally intended to have access to its Copyright Works) together with (ii) authorising reproduction by users. In this regard:
- as for the issue of communication to the public, Ms Lane correctly points out that Mr Saunders only raised concerns over this pleading for the first time on his feet, that there is a substantial amount of complex case law on communication to the public (none of which had been relied upon by the Defendant until Mr Saunders sought to rely in his oral reply upon an authority that had been included in the bundle for a different purpose) and that it has never previously been suggested by the Defendant that the Claimants cannot establish a real prospect of success in relation to communication to the public. No such case is advanced in respect of the allegations of primary infringement involving the text-to-image prompt, notwithstanding that exactly the same point (if correct) could presumably have been made on the SJ Application. In all the circumstances, I cannot conclude that the Claimants have no real prospect of success on this point. If this was a serious argument, I would have expected it to be identified well in advance of the hearing so that both sides could prepare to take the court to the relevant case law. As it is, the court has little more than a bare assertion from the Defendant that communication to the public cannot include the users with which this claim is concerned, notwithstanding its apparently inconsistent stance in not seeking to challenge the issue of communication to the public in connection with the text to image claim in its SJ Application.
- As for the issue of authorising reproduction by users, I did not understand Mr Saunders to suggest that this plea in itself failed to give rise to an arguable case.
Given that the Claimants say that every version of Stable Diffusion that is distributed to UK users is infringing, whether it is via Dream Studio or GitHub, I do not consider there to be any reason for present purposes to require particulars of the claim in so far as it concerns Dream Studio prior to granting permission. Particulars in relation to downloads via GitHub should suffice. Accordingly, the jurisdictional issue raised by the Defendant takes matters no further.