David Hernandez – Artificial intelligence (AI) is everywhere – in smartphones, cars, and even public places using facial recognition software. As companies rush to launch their own AI services, ChatGPT populates news cycles and laptops nationwide. But what exactly is AI? In short, it is the simulation of human intelligence by machines, and algorithms govern problem solving. Machine learning AI is made possible when computers learn from data and automatically adapt to information in a manner that mimics human activities.
Developments in AI have dispersed throughout the art world. “Edmond de Belamy,” an AI-generated portrait, sold at auction for $432,000. It was created by a Generative Adversarial Network (GAN), which utilizes differing, adversarial algorithms to improve its product quality. A generator model produces examples of images that do not exist, and a discriminator model determines whether those examples are real. When the discriminator fails to distinguish the real from fake, a more accurate image is created. Some AI models use diffusion, meaning that they aggregate keywords attached to millions of different images and merge those associated images together.
Despite AI’s technological intricacies, users can generate art by simply inputting words into the model. Through prompt engineering, one can easily manipulate the kinds of images the AI produces. Various combinations of nouns and adjectives, even requests to generate “in the style of x artist” all lead to countless possibilities. Users could further tailor their specific prompts with customized settings such as animation styles, aspect ratios, or definition quality. The very image shown above was created on StarryAI using the prompt “Artificial Intelligence Creates Art” alongside an Artstation style and photograph of Salvador Dali for reference. Other prominent examples of text-to-image AI include Midjourney, Stable Diffusion, and DALL-E 2.
Critics claim that prompt engineering is not a skill, and that AI-generated images are not actually art. According to Meriam-Webster, art is “the conscious use of skill and creative imagination especially in the production of aesthetic objects.” Photography was not initially considered art; it was commonly viewed as a “mechanism for replication” that offered no artistic quality. However, photographers utilize many techniques to capture creative images, from choosing different lenses to altering framing, lighting, and depth. Today, most can agree that photography is a credited art form. As such, AI art may be treated the same way. Each AI machine is already its own nuanced mode of creation, programmed with certain conditions and abilities that distinguish it from others. Furthermore, prompt-engineering could be analogized to a user’s paintbrush, and the AI platform to a canvas or camera.
Society may soon accept AI art, but the relevant laws paint a different picture. Under the Copyright Act, a work qualifies for protection if it’s (1) an original work of authorship, (2) fixed in a tangible medium, (3) that has a minimal amount of creativity. In February 2022, the U.S. Copyright Office (USCO) affirmed its rejection of Stephen Thaler’s copyright request for his AI-generated art piece. The application originally listed the author as a “Creativity Machine.” Thaler sought to register the work to the owner of the AI as a “work-for hire.” USCO ruled that the work needed human authorship to qualify, which was lacking because Thaler didn’t prove “sufficient creative input or intervention by a human Author.” USCO further argued that caselaw suggests that “non-human expression is ineligible for copyright protection.” The board interestingly referenced Burrow-Giles Lithographic Co. v. Sarony, where the Supreme Court ruled that photography was protected by copyright; the Court articulated that an author is “he to whom anything owes its origin,” and that photographs were “representatives of original intellectual conceptions of [an] author.” If photography is indeed comparable to AI art, then the courts and Congress should consider extending this rationale to AI art.
Advocates for the technology may support the existence of sufficient human intervention because customized prompt-engineering and AI settings result in unique, creative works. Other questions come to mind. What if users create portfolios of AI generated images that follow specific themes? What if AI models generate movies entirely via prompt-engineering? Moreover, artists could use AI as a means for expression. For instance, artists could upload their own pieces and then proceed to manipulate the prompt or settings to create new AI derivative works. The AI would act like a tool, similar to software used to make digital art. Artists would essentially be creating digital, original works that require at least minimal creativity. Through this lens, AI artwork ought to qualify for copyright protection, so long as AI isn’t listed as the author. USCO seems to agree; artist Kris Kashtanova received copyright registration for her graphic novel, which features content derived from Midjourney’s diffusion AI. Kashtanova stated that her work was AI-assisted, as she created the story and visual layout but used AI for artistic design. However, USCO is reconsidering the registration; the ambiguity highlights continued uncertainty.
Although copyright protection has benefits, the accessibility and relative ease of creating AI-art may lead to large waves of copyright requests, which could delay reviewing processes and heighten standards, causing more harm than good. There is also confusion as to who owns the AI output. For example, OpenAI’s terms for DALL-E 2 state that users are assigned the right, title, and interest in the output. But the terms also acknowledge that other users may generate similar images. Inversely, Midjourney refers to outputs created with their model as “assets,” and users grant Midjourney an “irrevocable copyright license” to those assets.
Currently, AI art is spurring litigation. Artists filed a class-action lawsuit against Stability AI and Midjourney, alleging copyright infringement. Owned by Stability AI, Stable Diffusion utilized billions of images from LAION-5B, a large-scale dataset designed for research purposes, to train its AI models. Most of the images come from sources like Flickr and Pinterest; another source is DevianArt, a site where artists share their work and pay for subscriptions. The plaintiffs argue that the AI platforms commercially benefitted from the artists’ work without compensating them or acquiring their consent. The class-action also accuses platforms like DevianArt of conspiring with Stable Diffusion; according to the complaint, DevianArt failed to provide artists with licenses or revenues from DreamUp, a product that feeds art to Stable Diffusion. Notwithstanding the promising nature of AI art, the increasing prevalence of the technology demands legal clarity on intellectual property concerns, whether it be from the bench or by legislation.
Note: This blog post was initially written and submitted in February, 2023.
