Controversies of creation: AI outputs, human work, and intellectual property

Most artists desire greater recognition so they can reach a larger audience and advance their careers. However, that is not quite the case for Polish digital illustrator Grzegorz Rutcowski, known for the fantasy landscapes that feature in games like Sony’s Horizon Forbidden West and Ubisoft’s Dungeons & Dragons. Generative artificial intelligence (AI) has expanded Rutcowski’s recognition, but not in ways that are beneficial to him. As of October 5, 2022, the artist’s name was used in image-generating prompts over 93,000 times on the AI platform Stable Diffusion. Rutkowski states, “What is terrifying is that those AI art generators are using everything that they can collect on the Internet without any consent, violating copyrights.”

Generative AI poses a challenge to artists and other creatives because AI models are trained on huge datasets that include countless copyrighted materials and the outputs do not include attribution to the artists or any compensation for the use of artists’ work. As a result, many artists, authors, and other creatives are standing up and mobilizing to stop generative AI firms from engaging in this kind of activity by raising awareness and bringing lawsuits against AI companies. While these lawsuits have not yet concluded, this collective action shows that creatives are seeking greater control over their content and wish to shape this new technology.

The challenge posed by generative AI is a result of how AI models are created and refined. Such models use computer algorithms to “train” themselves on enormous datasets, gradually enhancing their abilities to produce outputs that are similar to their training data. Such training data consist of huge datasets that contain a great deal of information freely available on the internet, including artists’ and authors’ original copyrighted and trademarked works.

AI models are not sentient and do not differentiate between copyrighted and non-copyrighted works, making it the responsibility of the model builders to build in protections for artists and other creatives. However, AI companies have not refined their models nor removed copyrighted works from the training datasets. This is partially because deleting billions of copyrighted works in training data would be very labor-intensive. Also, removal of these copyrighted materials, which are generally of higher quality, could undermine the model’s accuracy or the quality of its outputs. If AI companies are capable of furnishing credit and compensation to creatives their models take as references, it would not be a problem. However, it is not the case. The learning process used by AI models is extremely complex and often does not unfold in the same way each time, which means that neither the models themselves nor the companies know which artists’/writers’ copyrighted works their algorithms reference when producing outputs. This problem is worsened by the fact that most AI programs are either free or charge a small monthly fee, meaning that many people can produce potentially copyright-infringing outputs.

            There are three occupations whose interests are noticeably damaged by AI’s copyright violations and who are taking collective action to stop it: writers, artists, and music producers. One such action is the Writer’s Guild v. OpenAI case. The collective action began when a group of writers witnessed the ability of AI to duplicate their work. George R.R. Martin, who in 2023 was still working on the final pages of The Song of Ice and Fire series, found that a fan used ChatGPT to finish the series. Other authors, like Jane Friedman, found "a cache of garbage books" derived from her works on Amazon that she believes could only be AI-generated.

In order to put a stop to the unauthorized production of their work, the Author’s Guild, accompanied by authors like Martin, John Grisham, Michael Connelly, and Jodi Picoult collectively sued OpenAI. They argue that ChatGPT’s outputs do not fall under “fair use” (in which copyrighted works can be used without permission), as AI models are a commercial product whose owners profit from the authors’ original works. This argument is now undergoing a major test in the courts. In order to prove infringement, the plaintiffs deployed three lines of evidence: (1) generative AI is capable of returning relatively accurate quotations of copyrighted books, (2) the presence of AI-created derivative works suggests that the models had access to the copyrighted originals, and (3) one group of AI researchers states that books are essential for chatbot development. OpenAI, on the other hand, asserts that this is indeed a case of fair use since ChatGPT is a transformative work that conjures up new and creative understandings without replacing or challenging authors’ works in the marketplace.

In addition to the OpenAI vs Writer Guild case, in October 2023, visual artists filed suit in a California federal court targeting the AI company Stability for its use of copyrighted images without permission to create its AI program Stable Diffusion. In January 2024, the Recording Industry Association of America, accompanied by music industry titans, sued AI developers Udio and Suno for the company’s “mass infringement of copyrighted sound recordings.”

Though creatives are engaging in collective action in response to the challenge of AI-generated content, most of these cases face two obstacles: first, it is extremely difficult for creatives to determine which of their works have been copied—which would be solid evidence proving that companies actually use their works in training—because, according to how the AI training process works, none of the outputs are likely to be a close match for any specific copyrighted input.    Second, many AI companies maintain their actions are lawful by relying on the fair use doctrine. In order to win the first round, the content creators have to first find adequate solutions to these problems .

These collective actions give us a glimpse of the two possibilities for the future of AI development and the corresponding situations faced by content creators.On one hand, if the AI companies succeed in decisively proving that the usage of copyrighted materials in AI training is fair use, then copyrighted materials will continue to be used to train models, eventually resulting in even more powerful applications. However, this scenario won’t be a pleasant one for creatives because it increases the possibility that they (or at least parts of their work) will be replaced by AI. On the other hand, if creatives offer persuasive evidence that AI companies’ activities are not sheltered by fair use, the AI industry will have to look elsewhere for high quality training data. If such companies are not allowed to use existing training data, existing models would have to be rebuilt and re-trained on new data, a process that will probably be costly and time-consuming. AI is a new technology and is bringing with it new challenges and it will take time for humans to figure out how to coexist with it.