Generative AI and Artistic Rights

Evolving generative AI technologies have begun to transform the art industry. There are questions regarding the legality of the data that companies use to train these expansive AI models, and many artists have already experienced the negative impacts of AI.

This paper explores the legal, practical, and technological responses to these issues, emphasizing the need for transparent data practices and robust licensing agreements to protect artists’ rights and creativity.

Selected excerpts:

Development of Generative AI and Art
In 2020, a groundbreaking paper written by theoretical physicist, Jared Kaplan, highlighted that the more data used to train an AI model, the better it would perform. Prior to the publication of this paper, AI models were trained using relatively little training data. However, Kaplan’s paper made it clear that to perform efficiently and accurately, these models must be trained using extremely large datasets, often consisting of copyrighted materials scraped from the internet. Larger datasets resulted in AI models that could create detailed outputs faster than most humans could think. 

These technological advancements have revolutionized the industry at a cost to artists. Companies prioritize rapid development and outperforming competitors over ethical concerns and the rights of artists because it would “take too long” to negotiate licenses…

Preventing the Use of Data
Concerns over copyright infringement have led to the development of websites such as “Have I Been Trained,” where artists can search for their work in popular AI training datasets. The website allows users to find exact matches using image captions, artist names, or even a description of the image. A quick search on the site reveals that it uses everything from Van Gogh to elementary school artwork. The website also has a “Do Not Train Registry.” allowing users to claim their domain and set permissions on the usage of their images, with the goal of giving individuals more control over their creative work. This registry does not remove data from already trained models, nor is it binding upon AI companies, as they must agree to honor the do not train registry. The list of partner organizations that have agreed is growing and now includes HuggingFace (the largest repository of models and datasets) and Stability AI. 

The Glaze Project is a research effort that aims to develop technical tools to protect human creatives against infringing uses of generative AI. Its main goal is to disrupt unauthorized AI training on artists’ works and allow them to retain agency and control over the use of their work products. 

Style mimicry enables AI models to fine-tune images to the style of a specific artist. This can lead to the loss of commissions and income, as well as the dilution of their style, brand, and reputation that took years to develop. This has led to the demoralization of many young and aspiring artists, causing plummeting student enrollment. At Indiana University Bloomington, enrollment in the Arts and Sciences plummeted from 9066 students in 2014 to 7008 in 2022. Low enrollment has caused many art schools to shut down, including The University of the Arts, which was established in 1876. 

Glaze works to combat this by understanding the models that are training on human art and using machine learning algorithms to compute minimal changes to the artwork. These changes are nearly invisible to human eyes but create a significant difference in the output of an AI, protecting artists from style mimicry. It is nearly impossible to override Glaze because, unlike a watermark or steganography, it is a new dimension embedded within the artwork that a human cannot see, but an AI model can see, and is impossible to interrupt without an attacker knowing the specific dimensions…

Copyright Protections Before AI
Copyright laws protect original works that are independently created and sufficiently creative. “Creativity can be demonstrated in a variety of ways and reflects artistic choices like the subject matter, composition, depiction, and the use of the elements of design.” The Visual Artists Rights Act (“VARA”), a U.S. law, prevents others from “intentionally distorting, mutilating, or modifying artwork in a way that dishonors the artist’s reputation.” Although copyright protection exists from the moment the work is fixed, registering with the U.S. Copyright Office offers additional protections.

A derivative work is derived from one or more already existing works. Translation is considered derivative work, and to be copyrightable, it must incorporate some or all of the preexisting material. “A “compilation” is a work formed by the collection and assembling of preexisting materials or of data that are selected, coordinated, or arranged in such a way that the resulting work as a whole constitutes an original work of authorship.” Compilations of data and preexisting works are also copyrightable if the materials are selected or arranged in a way that the resulting work is considered a new work. Only the copyright owner of a work has the right to prepare or authorize another to create an adaptation of their work….

Read this, and so much more, including a summary of current litigation.