The Ethics of AI Art: Copyright Lawsuits Explained

The explosion of generative AI tools has revolutionized digital creation, but it has also sparked a fierce legal firestorm. If you are following the development of tools like Midjourney, Stable Diffusion, and DALL-E 3, you have likely heard about the “theft” allegations. Artists argue that these powerful models were built on the backs of their hard work without consent or compensation. This guide breaks down the complex legal battles currently unfolding in federal courts.

The Core Allegation: Scraping the Internet

To understand the lawsuits, you must first understand how these AI models learn. They do not conjure images out of thin air. Instead, they are trained on massive datasets containing billions of images and text descriptions scraped from the public internet.

The most prominent dataset is LAION-5B, which contains over 5.85 billion image-text pairs. This dataset includes everything from stock photos and medical images to copyrighted artwork found on platforms like ArtStation, DeviantArt, and Pinterest.

The central legal argument from the artist community is straightforward:

• Companies scraped copyrighted works to build a commercial product.
• They did not ask for permission.
• They did not offer compensation.
• The resulting product competes directly with the original artists in the marketplace.

Andersen v. Stability AI: The Class Action Suit

The most significant legal challenge in the United States is the class-action lawsuit Andersen et al. v. Stability AI Ltd. et al. Filed in the Northern District of California, this case targets Stability AI, Midjourney, and DeviantArt.

The Plaintiffs

Three prominent artists launched this battle:

• Sarah Andersen: A cartoonist known for “Sarah’s Scribbles.”
• Kelly McKernan: An independent artist whose style is frequently used as a prompt in AI generators.
• Karla Ortiz: A concept artist and illustrator who has worked for Marvel Studios.

The Legal Arguments

The plaintiffs argue that AI image generators are essentially “collage tools” that remix existing copyrighted works. They allege direct copyright infringement, violation of the Digital Millennium Copyright Act (DMCA), and violations of the right of publicity.

However, the defense argues that these models do not store copies of images. Instead, they argue the software analyzes mathematical patterns (pixel arrangements) to “learn” what a dog or a sunset looks like. They contend that generating a new image based on those learned patterns is not infringement, similar to how a human art student learns by studying the Masters in a museum.

Current Status of the Case

The road for the plaintiffs has been bumpy. In October 2023, Judge William Orrick dismissed many of the initial claims. He noted that the artists failed to provide specific registrations for every image they claimed was used. Furthermore, he was skeptical of the idea that every output from an AI model is automatically a “derivative work” of the training data.

However, the judge did allow the core claim of direct infringement against Stability AI to proceed. He also allowed the artists to amend their complaint to add more specific details and new plaintiffs, keeping the legal threat very much alive.

Getty Images vs. Stability AI

While the artists’ lawsuit focuses on style and livelihood, a separate lawsuit filed by Getty Images presents a more technical challenge for AI companies.

Getty Images sued Stability AI in both the US and the UK. Their argument is bolstered by concrete evidence: AI-generated images that display a distorted version of the famous Getty Images watermark.

This is a critical problem for the defense. It suggests that the AI model (Stable Diffusion) ingested Getty’s intellectual property so thoroughly that it memorized the trademarked logo. This case avoids the abstract debate about “artistic style” and focuses on the unauthorized processing of millions of proprietary stock photos. If Getty wins, it could force AI companies to pay substantial licensing fees for training data.

The "Fair Use" Defense

The entire future of generative AI hinges on two words: Fair Use.

Under US copyright law, you can use copyrighted material without permission if the use is considered “transformative.” This means the new work adds a different purpose or character to the original.

AI companies rely on the legal precedent set by Google Books. In that case, courts ruled that Google could scan millions of copyrighted books to create a searchable database because it was a “transformative” use (creating a search index) rather than a replacement for the books themselves.

AI lawyers argue that training a model to understand the relationship between pixels and words is highly transformative. They claim the output (a new image) is distinct from the input (the training data).

Artists counter that this is not transformative because the AI serves the exact same market purpose as the original art: creating visual media for entertainment or commercial use. If an art director uses Midjourney to generate a “Kelly McKernan-style” poster instead of hiring Kelly McKernan, the AI has acted as a direct market substitute.

Can You Copyright a Style?

A major hurdle for individual artists is that style is not copyrightable.

You can copyright a specific painting of a wizard on a mountain. You cannot copyright the general idea of painting wizards with thick brushstrokes and a specific color palette.

This distinction causes friction in the prompt engineering community. Users can type “in the style of Greg Rutkowski” (a popular fantasy artist) and generate work that mimics his aesthetic perfectly. While this feels like theft to the artist, it falls into a legal gray area. The courts must decide if an AI model mimicking a style constitutes infringement or if it is simply an advanced form of inspiration.

The Copyright Office Stance

While the lawsuits focus on the input (training data), there is also a battle over the output.

The US Copyright Office has firmly stated that images created entirely by AI cannot be copyrighted. In a ruling regarding an image created by the AI system “Creativity Machine,” the office declared that copyright requires “human authorship.”

This creates a paradox for businesses. A company might use Midjourney to generate marketing assets to save money, but they do not own the copyright to those assets. Competitors could technically copy those AI-generated marketing images without legal repercussion.

What Happens Next?

The legal landscape is shifting rapidly. Here are the specific developments to watch:

1. Amended Complaints: The artists in the Andersen case have filed amended complaints providing more evidence of how their specific works were used.
2. Discovery: As these cases move to the “discovery” phase, tech companies may be forced to reveal exactly what data is in their training sets, removing the current secrecy.
3. Opt-Out Mechanisms: Companies like Spawning have released tools like “Have I Been Trained?” which allow artists to see if their work is in the LAION dataset. Stability AI has begun offering opt-out processes, though critics argue this should be “opt-in” by default.

Frequently Asked Questions

Is it illegal to use Midjourney or Stable Diffusion? No. Using these tools is currently legal. The lawsuits target the companies building the tools, not the end-users creating images.

Do artists get paid when AI uses their work? Currently, no. There is no standard system for royalties or micropayments for training data. Some companies, like Adobe Firefly, are training on stock images they have rights to, ensuring contributors get paid, but this is the exception, not the rule.

Can I copyright an image I made with AI? Generally, no. The US Copyright Office has ruled that AI-generated content lacks the “human authorship” required for protection. However, if you significantly edit or paint over the AI image in Photoshop, you may be able to copyright the human-created changes.

How do I know if my art was used to train AI? You can use the website “Have I Been Trained?” to search the LAION-5B dataset. If your art appears there, it was likely used to train Stable Diffusion and possibly other models.