The strange thing is that Midjourney cannot understand fingers, considering that the generative AI has the knack for creating interesting images with little more than a simple text prompt. We can take, for example, the assembly line of stills from movies that were never made by famous directors or the more bizarre outputs from velociraptors posing with high school graduates at their proms or the live-action 1980s sitcom ‘King of the Hill’ reimagined with velociraptors. The digits, however, are not a particularly important part of these pictures, often having seven or eight of them to a hand or shaved at the knuckle and being overshadowed by more distracting and dazzling visuals.
Artists worldwide aren’t pleased with the finger crisis but are concerned about the possibility that generative platforms like Midjourney, DALLE-2, and Stable Diffusion may erode the integrity of their craft. However, some believe these models are little more than cheap remakes of the existing, copyrighted works being produced. Three of these artists filed a civil suit in San Francisco and London last month against Midjourney and Stability AI, alleging they were using their works in training databases without permission. As a result, any works produced by the models were derivative works. After several generations of photographs appeared containing ghostly reproductions of the popular watermark that is part of the platform’s famous watermark, stock photography provider Getty has written a letter warning both platforms of possible legal action in the future.
This kind of litigation, explains Cerys Wyn Davies, an intellectual property expert at Pinsent Masons, presents regulators with several challenges when balancing users’ interests with those of the industry they’re trying to transform through the use of new technology. According to Davies, you have to keep a balance between the rights owner and the freedom of people to innovate to maintain a healthy market. Among the main arguments involved in the lawsuit against Midjourney and Stability AI are that they violated the artists’ copyrights by using their images without permission and that including these images in the training data effectively creates derivative content as a consequence of the infringement of the artist’s copyright. According to Dr. Bahne Sievers of the University of California, Los Angeles, the argument about training data is perfectly reasonable. FieldFisher’s intellectual property lawyer explains that most jurisdictions do not allow anyone to use copyrighted data unless they have a license, which they can obtain after paying a fee. No one indeed has a license!”
A ‘fair use’ doctrine in the US, for example, permits using copyrighted materials to express freedom of expression or for ‘transformation’ under certain circumstances. It is, however, much more difficult to use copyrighted materials in European jurisdictions, as countries such as Germany and the UK are only willing to permit you to use copyrighted materials in certain circumstances. It would not be illegal to quote a copyrighted work or paint a portrait in a similar style to another, say, Sievers and Davies. Still, hoovering up artistic images to train an AI model without the author’s permission would not be unlawful.
“It is probably more difficult for Stability AI to argue for a broader fair dealing defense in the UK than it is in the United States, according to Gill Dennis, an intellectual property lawyer at Pinsent Masons, since they are not necessarily within any of the specific exceptions for fair dealing outlined in the legislation, and he believes this will make things much tougher for Stability AI.” It seems strange to some lawyers that generative AI companies are being sued, whereas those that compiled the dataset are not being sued.
In addition, Guadamuz says that it is also hard to prove that the works produced by generative AI are in some way reproductions of original works that were created by human beings. A photo, a painting, or a rendering is broken down into mathematical abstractions, which the model then uses to help chart the commonalities between the aspects of a particular photo, painting, or rendering and the written text prompts accompanying the image. To learn how to create something new, an artist might commit all of the pictures in the world to memory instead of incorporating fractional chunks of each image into their latest creation to learn how to create something new. However, the current wave of AI lawsuits doesn’t likely delay or halt the release of any new AI products soon. The future of the field is now contingent on players such as Midjourney, Stability AI, or even OpenAI since the general principles guiding the development of such generative applications have now been firmly established. They may get sued out of existence,” said Guadamuz, “and artificial intelligence would continue to develop at the same pace as it currently is.”
