The Colorado State Fair held its art competition in 2022. An AI-generated piece of artwork won. Jason Allen used Midjourney, a generative AI trained on art scraped off the internet, to create his work. Allen spent 80 hours and 900 iterations to refine his work.
His use of AI in winning the art competition has sparked a heated online backlash; one Twitter user claimed: “We are watching the death artistry unfold before our very eyes.”
Questions about authorship and ownership have also been raised as generative AI tools such as Midjourney and Stable Diffusion are brought to the forefront.
The AI is trained to generate art from a large number of previous artworks.
Should artists compensated for scraping their art to train models? Who owns images produced by AI systems? Is the process for fine-tuning prompts to generate AI a form of authentic creative expression?
technophiles adore Allen’s work. On the other hand, many working artist consider using their art to teach AI to be exploitation.
We are part of a group of 14 experts from different disciplines who have just published an article on generative AI for Science magazine. We explore the impact of AI on creative work, aesthetics, and media. One of the most important questions raised was about U.S. laws. Copyright laws and their ability to deal with the challenges of generative AI.
Copyright laws were designed to encourage the arts and creativity. The rise of generative AI, however, has complicated the notion of authorship.
The lens of photography is a useful tool.
While the concept of generative AI may seem new, history can serve as a useful guide.
Consider the rise of photography during the 1800s. Prior to the invention of photography, artists were limited to drawing, painting, or sculpture as a way to depict reality. A camera and chemicals allowed the capture of truth in an instant.
Many people argued, as with generative AI, that photography lacks artistic merit. In 1884, U.S. The Supreme Court of the United States weighed in and found that cameras were tools an artist could use to give an idea of visual form.
Since then, photography has evolved into a unique art form. It even inspired new artistic movements.
AI cannot own its outputs.
Unlike inanimate cameras, AI possesses capabilities – like the ability to convert basic instructions into impressive artistic works – that make it prone to anthropomorphization. The term “artificial intelligent” itself encourages people to believe that these systems are capable of humanlike intelligence or even self-awareness.
The U.S. Copyright Office has stated unequivocally that only humans can hold copyrights. The Copyright Office is clear that only humans are able to own copyrights.
Who can then claim ownership of AI-produced images? Who owns the photos that were used to train AI systems? Users who enter prompts in order to create images? Or those who make AI systems?
Is it fair use or infringement?
Generative AI uses data from training to produce results. Artists draw indirectly on past works to inspire and educate them.
These training data are a collection of artworks that have been collected without the consent or knowledge of artists. Many of these works are protected under copyright laws. This could violate copyright laws even before AI creates a new piece.
Memo Akten’s 2021 ‘All watched by machines of loving Grace.’ Created with custom AI software. Memo Akten, CC BY-SA
Midjourney, the artist who trained Jason Allen for his award-winning artwork, was familiar with 100,000,000 previous works.
Was it a form of infringement? Was it a form of ” Fair Use,” a legal doctrine that allows the unlicensed usage of protected works as long as they are sufficiently transformed into something else?
The AI systems may not have exact copies of the data, but they can recreate the data from time to time, which complicates the legal analysis.
Will the current copyright laws favor companies and end users over artists whose work is contained in training data?
Some scholars have proposed new regulations that would protect and compensate artists who use their work for AI training. These proposals include giving artists the right to refuse to have their data used to create generative AI or a way to automatically compensate artists if their work is used for training an AI.
Ownership confusion
However, training data is only one part of the entire process. Artists who use generative AI often go through several rounds of revision in order to improve their prompts. This suggests that they are original.
To answer the question of who owns the outputs, it is necessary to look at the contributions of everyone involved in the AI supply chain.
Legal analysis is made easier when the output differs from what was in the training data. In this instance, the AI appears to have defaulted to the owner of the production.
Copyright law, however, requires a meaningful creative contribution – which is not satisfied simply by pressing the shutter button of a camera. Uncertainty remains about how the courts will determine what this means in terms of using generative AI. Is the act of composing and refining a prompt sufficient?
It becomes more complex when the outputs look like works from the training data. It is unlikely that the copyright will be violated if there is only a general similarity in style or content. Fashion is not copyrightable.
Hollie Mengert experienced this problem first-hand when generative AI engines mimicked her style in a manner that, in her opinion, did not make her work unique. The singer Grimes, on the other hand, embraced this technology, “open-sourcing her voice” and encouraging fans to create songs using generative AI.
