Protecting Artists from AI: A Look at Glaze
With the rise of text-to-image models like MidJourney and Stable Diffusion, the art world has entered a new frontier — one that’s both exciting and unsettling. On one hand, AI is helping people create visually stunning pieces with just a few text prompts, but on the flip side, it’s threatening the livelihoods of professional artists. These AI models are learning how to mimic specific artistic styles after being trained, or “fine-tuned,” on samples of an artist’s work. Imagine pouring years into developing your own unique style, only for an algorithm to duplicate it in minutes. That’s where Glaze steps in. In a paper by Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y. Zhao from the Department of Computer Science ,University of Chicago, the researchers introduce Glaze, a tool designed to protect artists from AI style mimicry.
How does Glaze work? Basically, it allows artists to apply what they call “style cloaks” to their art before sharing it online. These cloaks make tiny, barely noticeable changes to the image. To the human eye, the artwork still looks the same, but when an AI model tries to learn from it, the model gets confused. The result? The AI fails to properly mimic the artist’s style, offering a safeguard for artists who want to protect their creative identity. Here’s the cool part: Glaze messes with the AI’s learning process without messing up the art itself. These tiny changes, or “perturbations,” don’t interfere with how people see the artwork, but they throw AI models off course when they try to copy it. When an AI model like Stable Diffusion attempts to learn an artist’s style from a Glaze-protected piece, it doesn’t pick up on the true stylistic nuances, making it much harder for the AI to recreate the artist’s work.
Now, it’s one thing to create a tool, but it’s another to get the people it’s meant for to actually use it. To see how artists feel about Glaze, the researchers conducted user studies with over 1,000 artists. These artists shared their thoughts on AI art, tested the tool, and gave feedback on its effectiveness, usability, and whether the changes to their art were tolerable. Spoiler alert: artists loved it. Most artists found Glaze to be effective in stopping AI from copying their style, even at low levels of perturbation (where the changes to the art are almost invisible). At a perturbation level of just p=0.05, the tool disrupted AI mimicry with over 92% success. Even when facing adaptive countermeasures — basically, AI models that try to outsmart Glaze — the tool still had an impressive success rate of over 85%.
For artists, this tool couldn’t come at a better time. The professional art community has been vocal about the risks AI poses to their careers. It’s one thing for AI to generate random images, but when it starts copying specific, well-established styles, that’s when things get murky. Glaze offers a layer of protection that allows artists to confidently share their work online without worrying about an AI hijacking their style.
The AI revolution in art isn’t slowing down, and tools like Glaze are going to be crucial in balancing innovation with protecting the integrity of the creative community. While AI models are only going to get smarter, Glaze proves that artists aren’t powerless in this new landscape. By using technology to fight technology, Glaze shows that the future of art can still have a human touch.
