How Artists Combat Copyright Infringement by Artificial Intelligence

What methods do artists have to combat the actions of technology companies that want to use their works to train artificial intelligence? A group of researchers has proposed a novel idea: adding trace toxins to artworks to defeat AI art generators from within.

Ben Zhao, a computer science professor at the University of Chicago, is a staunch critic of AI data analysis practices. He told MIT Technology Review that his team has developed a tool called "Nightshade" that lives up to its name - it poisons any model that uses images to train AI. Until now, artists' only options to fight AI companies were to sue them or hope that developers would comply with their requests to opt out.

This tool manipulates images at the pixel level to subtly distort them. Once enough of these distorted images are used to train AI, the entire model starts to break down. After introducing a data sample into a version of SDXL, the model interpreted prompts about "cars" as "cows." A dog was interpreted as a cat, and a hat turned into a cake. Similarly, different styles produced various bizarre deformations. Prompting for a "cartoon" resulted in artwork resembling 19th-century impressionism.

It can also be used to protect artists. If SDXL is used to create a painting in the style of renowned sci-fi and fantasy artist Michael Whelan, the poisoned model will produce works that are distinctly different from his.

Depending on the size of the AI model, hundreds or even thousands of poisoned images may be needed to generate these strange visual illusions. However, this may force anyone developing new AI art generators to think twice before using training data collected from the internet.

What tools do artists have to fight AI training?

Zhao is also the leader of the team that developed Glaze, a tool that creates a "style cloak" to confuse AI art generators attempting to mimic artists and their works. Zhao told MIT Technology Review that Nightshade will become another tool within Glaze, but it will also be released on the open market for other developers to create similar tools.

Other researchers have found ways to protect images from direct manipulation by AI, but these techniques do not prevent the use of data scraping techniques for training art generators. Nightshade is one of the most aggressive attempts to provide protection for artists so far.

Another emerging effort is trying to distinguish between real images and those generated by AI. DeepMind, a subsidiary of Google, claims to have developed a watermark that can identify whether an image was generated by AI, regardless of how it has been manipulated. These watermarks manipulate pixels in a way that is imperceptible to the human eye, similar to Nightshade. Some major AI companies have promised to watermark their generated content in the future, but no results have been seen yet.

Nightshade has the potential to be disruptive for companies that use artists' works to train their AI, such as DeviantArt. The DeviantArt community has had a fairly negative reaction to the AI art generator built into the website, and if enough users poison their images, developers will have to manually find every image that has been poisoned or retrain the entire model.

However, the program cannot change existing models like SDXL or the recently released DALL-3. These models have already been trained based on artists' past works. Companies like Stability AI, Midjourney, and DeviantArt have faced lawsuits from artists for using copyrighted works to train AI. Many other lawsuits have been filed against AI developers for using unauthorized copyrighted works, including Google, Meta, and Open AI. Companies and AI supporters argue that since the generated AI is creating new content based on training data, all the books, papers, images, and artworks in that training data fall under fair use.

Developers at Open AI have stated in their research paper that their latest art generator can create more realistic images because it was trained on detailed captions generated by the company's custom tools. The company has not disclosed the actual amount of data used to train its new AI model (most AI companies are reluctant to disclose any information about their AI training data), but as efforts to combat AI escalate over time, artists may take multiple measures to fight against them.