Since the rise of generative AI, many artists are worried about losing their livelihood to increasingly popular AI tools. There are many examples of companies replacing human work with such systems. Recently, Coca-Cola sparked controversy by creating a new Christmas advertisement using generative AI.
Visual artists, writers, and musicians have filed lawsuits against AI companies because their works were used in databases for training AI models without consent or compensation. Technology companies believe that anything on the public internet falls under the American Fair Use policy and can be used as long as the content is freely accessible. It may take years to find a legal solution to this problem.
Unfortunately, artists can do little if their work has already been included in a training dataset and used in a model that is already in circulation. However, they can take steps to prevent their work from being used in the future. Here are four ways how this can be done:
1. Masking the Style
One popular method for artists to protect against unwanted “AI scraping” is using masks on their images to protect their personal style from being copied. Tools like Mist, Anti-DreamBooth, or Glaze make tiny changes to the pixels of an image, invisible to the human eye, so machine learning models cannot properly “decode” the images if scraped. While Mist and Anti-DreamBooth require some programming knowledge, Glaze, developed by researchers at the University of Chicago, is easier to use. The tool is free and can be downloaded as an app or used online. Glaze is the most popular tool among the three, with millions of downloads. However, researchers from ETH Zurich and Deepmind demonstrated that “style cloning” can be relatively easily bypassed.
What works today might not work tomorrow. In computer security, bypassing such protection mechanisms is common among researchers as it helps find vulnerabilities and make systems safer. Using these tools is a calculated risk: once something is uploaded online, you ultimately lose control over it and cannot protect the images retroactively if the methods are insufficient.
2. Rethink Where and How to Share Images
Popular artist and photo sites like DeviantArt and Flickr have become gold mines for AI companies seeking training data. Once images are shared on platforms like Instagram, the parent company Meta can use the data to build their models if shared publicly, at least in the USA.
One way to prevent scraping is simply not to share images publicly online or keep social media profiles private. But for many creatives, that’s not an option as sharing work online is important for gaining clients. It might be worth considering sharing work on alternative platforms like Cara, a new project founded in response to artists’ criticism of AI. Cara, which collaborates with the researchers of Glaze, plans to integrate the researchers’ tools from Chicago. It also automatically implements “NoAI” tags, telling online scrapers not to scrape images from the site. Currently, the platform relies on the goodwill of AI companies to respect the artists’ wishes, but it’s better than nothing.
3. Opt-Out from Scraping
Using data protection laws, you can request technology companies to exclude your data from AI training. If you live in a country with such laws, like the UK or the EU, you can ask tech companies not to use your data for AI training. Unfortunately, opt-out requests from users in countries without such data protection laws are only considered at the tech companies’ discretion.
On the website “Have I Been Trained,” set up by the artist-run company Spawning AI, you can find out if your images have ended up in popular open-source AI training datasets. The organization has partnered with two companies: Stability AI, which developed the image generator Stable Diffusion, and Hugging Face, which promotes open access to AI. If your images have been added to Spawning AI’s “Do Not Train Registry,” these companies have agreed to remove your images from their training datasets before training new models. Again, this relies on the goodwill of AI companies and is not an industry-wide standard.
4. If All Else Fails, Use Some “Poison”
The researchers at the University of Chicago who developed Glaze also created Nightshade, a tool that makes invisible changes to pixels. Instead of just making it difficult for AI models to interpret images, it can (hopefully) “break” future iterations of these models, causing them to behave unpredictably. For example, images of dogs might turn into cats, and handbags into toasters. According to the researchers, relatively few modifications are needed to have an effect.
Nightshade can be downloaded as an app. In the future, the team hopes to combine Glaze and Nightshade, but for now, the two protection tools must be applied separately.
This article is by Melissa Heikkilä, an editor at the US edition of MIT Technology Review, reporting on developments in artificial intelligence.