AI databases steal private information but regulations prevent misuse

For most experts, the biggest questions concerning AI and copyright relate to the data used to train these models. Most systems are trained on huge amounts of content scraped from the web; be that text, code, or imagery. The training dataset for Stable Diffusion, for example — one of the biggest and most influential text-to-AI systems — contains billions of images scraped from hundreds of domains; everything from personal blogs hosted on WordPress and Blogspot to art platforms like DeviantArt and stock imagery sites like Shutterstock and Getty Images. Indeed, training datasets for generative AI are so vast that there’s a good chance you’re already in one (there’s even a website where you can check by uploading a picture or searching some text). The justification used by AI researchers, startups, and multibillion-dollar tech companies alike is that using these images is covered (in the US, at least) by fair use doctrine, which aims to encourage the use of copyright-protected work to promote freedom of expression. 

When deciding if something is fair use, there are a number of considerations, explains Daniel Gervais, a professor at Vanderbilt Law School who specializes in intellectual property law and has written extensively on how this intersects with AI. Two factors, though, have “much, much more prominence,” he says. “What’s the purpose or nature of the use and what’s the impact on the market.” In other words: does the use-case change the nature of the material in some way (usually described as a “transformative” use), and does it threaten the livelihood of the original creator by competing with their works?

Considering the onus placed on these factors, Gervais says “it is much more likely than not” that training systems on copyrighted data will be covered by fair use. But the same cannot necessarily be said for generating content. In other words: you can train an AI model using other people’s data, but what you do with that model might be infringing. Think of it as the difference between making fake money for a movie and trying to buy a car with it. 

Consider the same text-to-image AI model deployed in different scenarios. If the model is trained on many millions of images and used to generate novel pictures, it’s extremely unlikely that this constitutes copyright infringement. The training data has been transformed in the process, and the output does not threaten the market for the original art. But, if you fine-tune that model on 100 pictures by a specific artist and generate pictures that match their style, an unhappy artist would have a much stronger case against you. 

“If you give an AI 10 Stephen King novels and say, ‘Produce a Stephen King novel,’ then you’re directly competing with Stephen King. Would that be fair use? Probably not,” says Gervais.

Crucially, though, between these two poles of fair and unfair use, there are countless scenarios in which input, purpose, and output are all balanced differently and could sway any legal ruling one way or another. 

Ryan Khurana, chief of staff at generative AI company Wombo, says most companies selling these services are aware of these differences. “Intentionally using prompts that draw on copyrighted works to generate an output [...] violates the terms of service of every major player,” he told The Verge over email. But, he adds, “enforcement is difficult,” and companies are more interested in “coming up with ways to prevent using models in copyright violating ways [...] than limiting training data.” This is particularly true for open-source text-to-image models like Stable Diffusion, which can be trained and used with zero oversight or filters. The company might have covered its back, but it could also be facilitating copyright-infringing uses.

Another variable in judging fair use is whether or not the training data and model have been created by academic researchers and nonprofits. This generally strengthens fair use defenses and startups know this. So, for example, Stability AI, the company that distributes Stable Diffusion, didn’t directly collect the model’s training data or train the models behind the software. Instead, it funded and coordinated this work by academics and the Stable Diffusion model is licensed by a German university. This lets Stability AI turn the model into a commercial service (DreamStudio) while keeping legal distance from its creation.

Baio has dubbed this practice “AI data laundering.” He notes that this method has been used before with the creation of facial recognition AI software, and points to the case of MegaFace, a dataset compiled by researchers from the University of Washington by scraping photos from Flickr. “The academic researchers took the data, laundered it, and it was used by commercial companies,” says Baio. Now, he says, this data — including millions of personal pictures — is in the hands of “[facial recognition firm] Clearview AI and law enforcement and the Chinese government.” Such a tried-and-tested laundering process will likely help shield the creators of generative AI models from liability as well.

There’s a last twist to all this, though, as Gervais notes that the current interpretation of fair use may actually change in the coming months due to a pending Supreme Court case involving Andy Warhol and Prince. The case involves Warhol’s use of photographs of Prince to create artwork. Was this fair use, or is it copyright infringement?

“The Supreme Court doesn’t do fair use very often, so when they do, they usually do something major. I think they’re going to do the same here,” says Gervais. “And to say anything is settled law while waiting for the Supreme Court to change the law is risky.”