How AI, like DALL-E, ChatGPT, Stable Diffusion are affecting copyright owners and creators

Can AI own copyright?

The first, and perhaps more settled, question is: can AI own copyright? The answer is fairly straightforward as far as the DMCA is concerned.

As the World Intellectual Property Organization (WIPO) explains, the current position in US law is that AI cannot own copyright. This stems from settled case law which specifies that the protection offered by copyright applies only to “the fruits of intellectual labor… founded in the creative powers of the [human] mind.” The US Copyright Office has also taken a similar position historically. Their previously published guidance indicates that the USCO will “register an original work of authorship, provided that the work was created by a human being.”

Emphasizing this position, the USCO in September 2022 rejected a copyright application for a comic book generated using Midjourney, a text-to-image AI. While the application was initially approved, it’s notable that the registration didn’t go to the AI, but to the artist who created the work. Elaborating on the approval, the artist stated that the USCO asked her to provide evidence “that there was substantial human involvement” in the creative process.

So, the bottom line is that AI cannot own copyright under the DMCA, at least for now. But humans can, if they can prove they had substantial input in creating the work.

What amounts to substantial input?

That becomes a question of degree. For instance, expending considerable effort honing prompts, coaching the AI towards the desired output, and editing the final work suggests a higher level of involvement than generating entire books using 50-word prompts on ChatGPT.

Ultimately, it boils down to how much of a role you played in the process and the degree of effort put into generating the final work.

What happens if you cannot show substantial input?

Unfortunately, the implication could be that the work won’t qualify for copyright. Keep in mind that a central theme of copyright law is that there must have been a creation process solely involving or substantially directed by a human. Without the input of a human, the conclusion (by the USCO or a court of law) could simply be that no one owns the copyright.

Notably, the UK pursues a different approach to the DMCA. It is one of few nations globally that allow copyright protection for solely AI-generated works. But the copyright won’t be going to the AI here. Instead, the UK deems the author to be the creator of the work or that person “by whom the arrangements necessary for the creation of the work are undertaken.”

Of course, whether the copyright holder in this case would be the developer of the AI or its operator is a different argument entirely. But at least there’s less chance of the work going without copyright in the UK.

That said, let’s consider the other side of the coin: if AI cannot own copyright, can it encroach on others’ copyright?

Can AI infringe on copyright?

There are two potential situations in which AI could violate someone else’s copyright. The first is in the case of AI training systems.

Infringement by AI training datasets

Typically, AI developers and researchers spend years and thousands of hours training AI systems with huge datasets. For example, the training data for Stable Diffusion, a leading text-to-image system, contained billions of images scraped from hundreds of websites across the internet.

With the outsized nature of these training datasets, there’s a risk that the material includes copyright-protected content. So, in such cases, it’s possible that AI might infringe on someone else’s copyright. However, on the flip side, there’s a possible justification for such use of copyrighted material based on the fair use doctrine.

The fair use doctrine allows the use of copyrighted works when the unlicensed use does not unfairly infringe on the copyright owner’s rights. Some factors it considers are whether the use of the material was transformative, how much of the protected work was used, and whether it threatens the business of the original creator by competing with their work.

The courts have held in favor of similar use in a number of cases, including popular judgments involving Google’s Book Search tool and the plagiarism detection software Turnitin. In the Turnitin case, the US court held that scanning student papers (which are protected by copyright) to detect potential plagiarism was fair use. Likewise, in Google’s case, the court decided that scanning books and providing snippets to web searchers amounted to fair use, because Google only did that to make the books more searchable and not give away their expressive content.

So, seeing as the goal of generative AI training sets is only to train the program to recognize certain patterns on its way to creating original works, it’s likely to be covered by fair use. Note that the fair use doctrine is currently before the Supreme Court in a lawsuit involving Prince and Andy Warhol, so elements of the principle might change soon.

But it’s a different story when it comes to using AI to generate potentially infringing content.

Infringement by AI-generated content

While training AI systems with copyrighted material might be allowable, using the trained model in a way that encroaches on copyrighted works likely isn’t. As The Verge aptly notes, it’s “the difference between making fake money for a movie and trying to buy a car with it.”

There are many scenarios in which copyright infringement may occur. For example, if you plugged several works by Rembrandt into an AI and asked it to paint you a new Rembrandt, that’s probably going to qualify as infringement. Even if the AI generates a Rembrandt that’s never been seen before, you’re directly competing with the artist and unfairly appropriating his unique style.

Likewise, using AI prompts to generate works that include elements of works owned by someone else would amount to infringement. For instance, it’s probably illegal to prompt DALL-E to create a new Superman comic for you. While the comic would probably be unique, you’re drawing on a character that is protected by copyright without permission, and that’s illegal.

Therefore, it’s important to be extra careful when generating any content with AI (whether music, art, text, or code) to ensure you’re not violating any copyright.

Keep in mind that most of these issues are as yet untested as it relates to generative AI. There’s an argument that AI training data should not be allowed to violate the copyright of others, especially when the AI is commercial in nature. While murmurings around these issues have been around for some time, the voices of dissent are likely to grow louder in the coming years.

Trouble already brewing

A lawsuit was recently filed against three prominent generative AI companies (DeviantArt, Stability AI, and Midjourney), alleging that they contributed to the “unprecedented open-source software piracy” enabled by the Stable Diffusion AI.

The lawsuit alleges that the companies violated the intellectual property rights of thousands of artists with their text-to-image systems with claims including right of publicity violations, DMCA breaches, and unlawful competition.

The entire industry will be watching this case keenly, as any decision there will have significant ramifications for the use of generative AI going forward.

In the meantime, it’s probably best to be a bit more circumspect about the use of generative AI programs in creating original content.