Legal AI Case Update
Tracking The Image Creators' Battle
As noted in one of our earliest posts, there are currently several lawsuits pending by various creative folks suing multiple AI image creation companies. The crux of the lawsuits are two primary concerns: 1 - Is it copyright infringement for AI companies to train their image creation models using copyrighted or otherwise protected works? 2 - Is the output of AI image creation tools, depending on its visual similarity to copyrighted works, also infringing copyrights when it produces images in response to prompts.
The AI related defendants have recently filed a motion to dismiss the entire lawsuit. Let’s dive in.
The opening factual claim which has not been reported is an attempt to educate the judge about what it means to “train” an AI model using images. The Defendants here make the point that “Stable Diffusion was trained on billions of images that were publicly available on the Internet. To be clear, training a model does not mean copying or memorizing images for later distribution. Indeed, Stable Diffusion does not ‘store’ any images.” Motion to Dismiss at p.1. Having spent years in courtrooms and on the CLE circuit nationwide teaching lawyers and judges about the fundamentals of technology, I can see that this nugget is designed to ensure no mistake here by the court. Using words like “training a model” does not convey anything to the non-techie. It is entirely possible a court reading that claim would think that the “model” now has in its possession a copy of all those billions of images.
The motion continues by explaining that the model training took those images and learned “what things look like.” Id. A smart point that followed is important. If the goal of the AI model training was merely to duplicate the artists’ works, there are far easier and cheaper ways to do that which have been in widespread use for decades. (Photoshop anyone?).
That leaves the court with the question, “well, what did this thing do then as part of the creation of that model?” A great response for the plaintiffs, however, is that the model could not have been created without data, their data, their works, which were required so that it could produce works based upon what those artists decided “things look like” in their original works. The two edged sword there with that argument. Continuing….
The defense highlights next an admission in the complaint. “[The] Complaint concedes, ‘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.’ (Compl. ¶ 93.)”. Uh, well, then. The model was trained on our stuff, the argument goes, but we agree that it’s output will not likely violate any copyrights on our stuff. I have no idea why the lawyers for the Plaintiffs felt the need to make such an admission in the Complaint. Having worked with two of these tools for more than a year, (and being a past court qualified digital imaging expert), the statement that no output could possibly infringe any artists work is unlikely.
Not being a copyright lawyer, I was unaware that a pre-condition to filing a copyright infringement action is that the plaintiff has to have actually registered the work with the copyright office pre-suit. Neither plaintiff did that. One of them registered 16 works after the suit was filed. Motion to Dismiss at p.2. They then go on to cite an impressive list of federal cases (including the U.S. Supreme Court) noting that merely amending a complaint filed pre-registration is not going to cut it. Seems like a problem there for the Plaintiffs even before getting to the meat of their claims. But, in reality, the suit could fold tomorrow and undoubtedly some other plaintiffs, who will register their copyrights before suit, can refile essentially the same claims and get past this hurdle.
The Defendants then point out, accepting the complaint allegations as true, that it fails to cite a single work by plaintiff that, in part, appears in an image produced by the Defendant’s AI tools. Id.
Infringement requires that the infringing work represent “substantial similarity” with the copyrighted work. Unfortunately for these Plaintiffs, their complaint admits that they have not found any output of the Defendants’ AI models that meet that standard. They do allege that “every work” from those models is “copied or derived” from Plaintiffs works. Id. at 6-7. Defendants cite a Ninth Circuit case in which that argument was rejected.
The Plaintiffs also made a vicarious infringement claim. Defendants point out that facts supporting several of the elements of that claim are not in the complaint. For example, the complaint, they argue, does not allege that the Defendants have made any income from infringing images created with their AI tool. Id.
The Plaintiffs also made a claim under the Digital Millenium Copyright Act (DMCA). That act prohibits the removal of features of a copyrighted work designed to thwart copying. (Napster just lives on and on in the law doesn’t it?).
One of the claims that seems novel and potentially viable is that the AI model enables the creation of art that is “in the style of” the artist. This represents an interesting consideration of copyright given that many of us can imagine an image of the skyline of our city, for example, “in the style of Van Gogh’s Starry Night.” These kinds of “in the style of” works are all over the Internet now. Some represent meaningful extensions and re-interpretations of the original artist’s style which they evoke and others are wholly new things, but undoubtedly inspired by that artist. Defendants point out, however, that the law does not find infringement for works “in the style of” another artist. Beyond the legal issue here, it is arguably the end of art as we know it if artists cannot draw on earlier works to some small or great extent. Hasn’t that always happened since the dawn of art? I am sympathetic to this claim from a creative perspective. Many people spend years developing, nurturing and commercializing a “style” that distinguishes them. Then, AI comes along and in minutes, works just like those of that artist are for sale across the Internet and the artist’s works become instantly devalued presumably. Years ago I had a friend who was contacted by a hotel chain interested in purchasing the right to reproduce her works for decoration of hotel rooms in a new series of buildings being constructed. That kind of income and work for artists today is now gone with AI. Organizations can simply prompt a model for new works “in the style of” Van Gogh, Rothko, Dali and duplicate those endlessly across their hotels or whatever. Alas, the law does not always track what we as individuals might think is “right.”
Will AI Developments Change The Outcome?
As many of you have no doubt seen, AI music is having a moment. A short audio file sounding just like Drake and an artist named The Weeknd recently went viral online. The entire song was the output of an AI model trained on the singing voices of the two artists. The music publisher for one of the artists sent out a raft of letters to get it taken down. It wasn’t fast enough to prevent millions of people from downloading it and many “reaction” videos online of people being told it was not the real artists as they bopped along to the tune.
If AI developments move as fast as many fear (hope) they will, by the time this digital imaging AI tool case reaches the U.S. Supreme Court perhaps the world of AI will frighten the Justices into a decision they might not otherwise have made. It is presumed that more liberal Justices would be inclined to be sympathetic to the Plaintiffs in this case. But, what of a conservative Justice? Can you imagine the memes across the Internet involving all of the Justices by now? And, for years going. Are any of them just fed up with the visual, audio and video content associated with them, some no doubt profane, others humorous and still others insinuating things about their character or that of their family members/friends? Will they be so sympathetic to a company that makes a tool that enables people to make such content with even greater ease? One wonders.