midjourney – Techdirt (original) (raw)

Judge Dismisses Most Of The First Of The Many ‘How Dare AI Train On My Material’ Lawsuits

from the did-chatgpt-write-the-original-complaint? dept

There are now a bunch of these lawsuits accusing AI companies of some sort of copyright infringement for training their models on works of plaintiffs. However, the first high profile one was the case brought by Sarah Andersen, Kelly McKernan, and Karla Ortiz against Stability AI, MidJourney, and (bizarrely) DeviantArt. We covered the case back in April when the companies moved to have the case dismissed.

As we noted at the time, somewhat incredibly, the artists hadn’t even registered the copyright in their their works before suing, which is the kind of error that dooms cases. It’s also embarrassing for their lawyers, Joseph Saveri and Matthew Butterick, who have been behind a bunch of these lawsuits, and present themselves as knowledgeable lawyers, yet couldn’t even think to make sure the works they sued over were registered? (For the non-copyright experts among you, you don’t need to register your works to have a copyright, but you do need to register them if you intend to sue).

Now the judge, William Orrick, had mostly dismissed the case, as we expected after the oral arguments made it pretty clear how weak this case was.

The first issue was the lack of registrations:

Each defendant argues that McKernan and Ortiz’s copyright claims must be dismissed because neither of them has registered their images with the Copyright Office. They also move to “limit” Anderson’s copyright claim to infringement based only on the 16 collections of works that she has registered….

In opposition, plaintiffs do not address, much less contest, McKernan or Ortiz’s asserted inability to pursue Copyright Act claims. At oral argument, plaintiffs’ counsel clarified that they are not asserting copyright claims on behalf of these two plaintiffs. July 19, 2023 Transcript (Tr.), pg. 17:1-5. As such, McKernan and Ortiz’s copyright act claims are DISMISSED WITH PREJUDICE.

Likewise, plaintiffs do not address or dispute that Anderson’s copyright claims should be limited to the collections Anderson has registered. The scope of Anderson’s Copyright Act claims are limited to the collections which she has registered.

Even in the limited case of Andersen, who had registered some works, the defendant companies pointed out that no actual arguments were made regarding which actual works the AI systems were trained on, but the judge says that part can at least move forward with discovery.

As we had noted earlier, the inclusion of DeviantArt never quite made any sense at all in this lawsuit, and again seemed to show that Saveri and Butterick had no clue what they were doing. DeviantArt did not create an AI system, but instead, the argument was that Stability’s AI was trained on DeviantArt works. As the court notes, that doesn’t make DeviantArt liable for any direct infringement:

Plaintiffs fail to allege specific plausible facts that DeviantArt played any affirmative role in the scraping and using of Anderson’s and other’s registered works to create the Training Images. The Complaint, instead, admits that the scraping and creation of Training Images was done by LAION at the direction of Stability and that Stability used the Training Images to train Stable Diffusion…. What DeviantArt is specifically alleged to have done is be a primary “source” for the “LAION-Aesthetic dataset” created to train Stable Diffusion…. That, however, does not support a claim of direct copyright infringement by DeviantArt itself.

The court does allow the plaintiffs to amend their complaint to argue that “compressed images” of the infringing works are somehow included in the AI training database, but seems skeptical.

There was another theory presented that because users of DeviantArt can use the DreamUp tool that is based on Stability that it is creating infringing works, but that’s not how any of this works. DeviantArt notes that no matter what prompt you put into the system, you’re not going to get “substantially similar” works out of the system, and they need to be substantially similar to infringe.

The plaintiffs argue that you don’t need substantial similarity because the new works are inherently derivative of the works it was trained on, but the court says that’s wrong:

Plaintiffs rely on that line of cases and point to their allegation that all elements of plaintiff Anderson’s copyrighted works (and the copyrighted works of all others in the purported class) were copied wholesale as Training Images and therefore the Output Images are necessarily derivative….

A problem for plaintiffs is that unlike in Range Road – observed wholesale copying and performing – the theory regarding compressed copies and DeviantArt’s copying need to be clarified and adequately supported by plausible facts. See supra. The other 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.

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”).

Then there’s MidJourney. The complaint argues that MidJourney uses Stable Diffusion, but MidJourney hasn’t actually said that anywhere, so the judge asks for clarification:

Plaintiffs need 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 vicarious infringement claims are all dismissed as well, with leave to amend, specifically regarding the claims that Stable Diffusion supposedly has “compressed copies” of the images it was trained on:

Plaintiffs have been given leave to amend to clarify their theory and add plausible facts regarding “compressed copies” in Stable Diffusion and how those copies are present (in a manner that violates the rights protected by the Copyright Act) in or invoked by the DreamStudio, DreamUp, and Midjourney products offered to third parties. That same clarity and plausible allegations must be offered to potentially hold Stability vicariously liable for the use of its product, DreamStudio, by third parties.

There’s also a claim over DMCA 1202(b), alleging that the defendants removed copyright information from the images, but as the defendants note, there are no allegations of any actual removal.

In response, plaintiffs point to paragraphs180 and 191 of their Complaint, where they allege generally that plaintiffs and “others” in the putative class included various categories of CMI in their works and the “removal or alteration” of that CMI by defendants, including “the creator’s name” and “the form of artist’s signatures.” These allegations are wholly conclusory. In order to state this claim, each plaintiff must identify the exact type of CMI included in their online works that were online and that they have a good faith belief were scraped into the LAION datasets or other datasets used to train Stable Diffusion. At the hearing, plaintiffs argued that it is key for the development of generative AI models to capture not only images but any accompanying text because that accompanying text is necessary to the models’ ability to “train” on key words associated with those images. Tr. at 9:13-24. But there is nothing in the Complaint about text CMI present in the images the named plaintiffs included with their online images that they contend was stripped or altered in violation of the DMCA during the training of Stable Diffusion or the use of the end-products. Plaintiffs must, on amendment, identify the particular types of their CMI from their works that they believe were removed or altered.

The publicity rights claims also go nowhere:

The problem for plaintiffs is that nowhere in the Complaint have they provided any facts specific to the three named plaintiffs to plausibly allege that any defendant has used a named plaintiff’s name to advertise, sell, or solicit purchase of DreamStudio, DreamUp or the Midjourney product. Nor are there any allegations regarding how use of these plaintiffs’ names in the products’ text prompts would produce an “AI-generated image similar enough that people familiar with Plaintiffs’ artistic style could believe that Plaintiffs created the image,” and result in plausible harm to their goodwill associated with their names, in light of the arguably contradictory allegation that none of the Output Images are likely to be a “close match” for any of the Training Images… Plaintiffs need to clarify their right of publicity theories as well as allege plausible facts in support regarding each defendants’ use of each plaintiffs’ name in connection with advertising specifically and any other commercial interests of defendants.

There were also unfair competition claims, but those are pre-empted by the Copyright Act (basically, the UCL claims are really an attempt to argue infringement under a different law, and you can’t do that).

DeviantArt also brought an anti-SLAPP motion, which the judge says he’ll take up after the plaintiffs present an amended version.

So, the end result is that only the claim by Andersen (misspelled as Anderson throughout the ruling) of direct infringement against Stability AI can move forward for its apparent collection of images for its training database. Even that may still end up on the losing end after discovery. But this at least allows for discovery to happen first. Basically all the other claims are dismissed, though the plaintiffs can still amend those, but they’re going to have to overcome some pretty massive inherent weaknesses.

The lawyers for the plaintiffs are claiming victory because of the one claim being allowed to go to discovery, but this is a pretty embarrassing outcome for them overall.

Filed Under: ai, copyrigght, copyright registration, direct infringement, joseph saveri, karla ortiz, kelly mckernan, laion database, matthew butterick, saraha andersen, training, william orrick
Companies: deviantart, midjourney, stability ai

from the copAIright dept

Over the last few months there have been a flurry of lawsuits against AI companies, with most of them being focused on copyright claims. The site ChatGPTIsEatingTheWorld has been tracking all the lawsuits, which currently lists 11 lawsuits, seven of which are copyright claims. Five of those are from the same lawyers: Joseph Saveri and Matthew Butterick, who seem to want to corner the market on “suing AI companies for copyright.”

We already covered just how bad their two separate (though they’re currently trying to combine them, and no one can explain to me why it made sense to file them separately in the first place) lawsuits on behalf of authors are, as they show little understanding of how copyright actually works. But their original lawsuit against Stability AI, MidJourney, and DeviantArt was even worse, as we noted back in April. As we said at the time, they don’t allege a single act of infringement, but rather make vague statements about how what these AI tools are doing must be infringing.

(Also, the lawyers seemed to totally misunderstand what DeviantArt was doing, in that it was using open source tools to better enable DeviantArt artists to prevent their works from being used as inspiration in AI systems, and claimed that was infringing… but that’s a different issue).

It appears that the judge overseeing that lawsuit has noticed just how weak the claims are. Though we don’t have a written opinion yet, Reuters reports that Judge William Orrick was pretty clear at least week’s hearing that the case, as currently argued, has no chance.

U.S. District Judge William Orrick said during a hearing in San Francisco on Wednesday that he was inclined to dismiss most of a lawsuit brought by a group of artists against generative artificial intelligence companies, though he would allow them to file a new complaint.

Orrick said that the artists should more clearly state and differentiate their claims against Stability AI, Midjourney and DeviantArt, and that they should be able to “provide more facts” about the alleged copyright infringement because they have access to Stability’s relevant source code.

“Otherwise, it seems implausible that their works are involved,” Orrick said, noting that the systems have been trained on “five billion compressed images.”

Again, the theory of the lawsuit seemed to be that AI companies cut up little pieces of the content they train on and create a “collage” in response. Except, that’s not at all how it works. And since the complaint can’t show any specific work that has been infringed on by the output, the case seems like a loser. And it’s good the judge sees that.

He also recognizes that merely being inspired by someone else’s art doesn’t make the new art infringing:

“I don’t think the claim regarding output images is plausible at the moment, because there’s no substantial similarity” between images created by the artists and the AI systems, Orrick said.

It seems likely that Saveri and crew will file an amended complaint to try to more competently make this argument, but since the underlying technology doesn’t fundamentally do what the lawsuit pretends it does, it’s difficult to see how it can succeed.

But, of course, this is copyright, and copyright caselaw doesn’t always follow logic or what the law itself says. So it’s no surprise that Saveri and Butterick are trying multiple lawsuits with these theories. They might just find a judge confused enough to buy it.

Filed Under: ai, copyright, direct infringement, inspiration, joseph saveri, matthew butterick, william orrick
Companies: deviantart, midjourney, stability ai

Midjourney CEO Says ‘Political Satire In China Is Pretty Not Okay,’ But Apparently Silencing Satire About Xi Jinping Is Pretty Okay

from the pretty-not-okay dept

As a rule, it’s a good idea to be particularly suspicious of defenses of censorship that — coincidentally — materially benefit the people espousing them. In this case, the argument in favor of censorship is coming from founder and CEO of AI image generator Midjourney, David Holz. And Holz makes clear that he is willing to exempt Xi Jinping from the tool’s capabilities to retain Midjourney’s viability in China.

That’s right: Xi Jinping, one of the most powerful and repressive government officials in the world and most deserving of political skewering and mockery will be one of the few exempt from it, at least where Midjourney is concerned. Some other terms are restricted, though Holz won’t make the list public — “Afghanistan” for example, and now some depictions of arrests after the fake Donald Trump arrest fiasco — but Holz reportedly treats China as a unique case.

His quote about it is a doozy. From The Washington Post:

But the year-old company, run out of San Francisco with only a small collection of advisers and engineers, also has unchecked authority to determine how those powers are used. It allows, for example, users to generate images of President Biden, Vladimir Putin of Russia and other world leaders — but not China’s president, Xi Jinping.

“We just want to minimize drama,” the company’s founder and CEO, David Holz, said last year in a post on the chat service Discord. “Political satire in china is pretty not-okay,” he added, and “the ability for people in China to use this tech is more important than your ability to generate satire.”

He wants you to simultaneously believe that his program is so important that it must do whatever is necessary to remain accessible to people within China, but so unimportant that it doesn’t matter if fundamental political expression about one of the most powerful authoritarians in the world can’t be created on it. It doesn’t add up.

It’s no surprise that a tech CEO would be willing to make trade-offs for the Chinese market. At this point, it’s more surprising if one won’t do so. But Holz’s position is particularly careless and reveals an increasingly serious threat to free expression on and offline today: individual countries’ censorship laws, particularly those of powerful countries like China, are setting global rules sometimes enforced by tech companies anxious to display their compliance. It’s not just Midjourney’s China-based users that can’t satirize Xi Jinping — that rule applies to users everywhere, even in the United States.

Local laws are suddenly not so local anymore, and people like Holz have no qualms about aiding their illiberal international spread.

What this means in practice is that authoritarian leaders don’t just get to subject their own countries to repressive laws limiting political speech. They also get to set the rules for global communities which are not, and should not be, under any expectations to abide by them. Zoom engaged in such practices in 2020 when it applied Chinese law to users outside mainland-China, shutting down online Tiananmen memorials held by users in Hong Kong and the United States. In response to well-deserved criticism, Zoom announced it would no longer allow Chinese law to dictate policies outside mainland China. Midjourney took notes, it seems, and learned a different lesson.

Holz’s exact words were “the ability for people in China to use this tech is more important than your ability to generate satire,” but his meaning was clear: Midjourney’s interest in being accessible in China is more important than its users’ interest in engaging in political expression.

This acquiescence signals to authoritarians of all sorts that if they want to control their image on the global internet, ramping up repressive efforts at home will be rewarded. As if they needed more incentive. Oversee a large enough financial market and censor enough people and you, too, may be able to control your reputation on the global internet.

Will Midjourney stop with just Xi Jinping? Or should we expect satire of Prime Minister Narendra Modi, for example, to be next on the chopping block if Midjourney wants access to an increasingly unfree India?

Lastly, Holz may argue that he’s helping Chinese citizens — and not just his own company — by attempting to ensure that Midjourney will be available in China. But his claim that he wants to preserve “the ability for people in China” to use this tech is paternalistic in a way Holz doesn’t seem to realize. Chinese citizens, after all, may themselves want to use Midjourney to satirize their own government. Why does their freedom to satirize their ruler matter less than the freedom of the rest of the world to mock their leaders?

In fact, in the later months of 2022, protesters in cities across China held up blank sheets of paper in country-wide demonstrations to protest not just the country’s restrictive COVID policies, but the many things that Chinese citizens aren’t allowed to voice. This censorship is a result of oppressive and wide-ranging governmental control over what they can say in every forum, online and off. But it’s reinforced when foreign companies, in this case a U.S.-based AI image generating tool, are all too eager to do their part in enforcing those restrictions — not just on the population legally bound by them, but on the rest of the world, too.

Companies like Midjourney may be the vanguard of new technology and the changing internet. But censorship is nothing new, and they won’t change the game by willingly conducting reputation management for authoritarian governments.

That, one might say, is “pretty not-okay.”

Sarah McLaughlin is Senior Scholar, Global Expression at the Foundation for Individual Rights and Expression.

Filed Under: ai, censorship, china, david holz, free speech, generative ai, political speech, satire, xi jinping
Companies: midjourney