A new proof-of-concept technique called LightShed is testing the strength of one of artists’ most visible technical defenses against AI training. The tool is designed to remove the subtle image changes that programs such as Glaze and Nightshade add to digital art in an effort to disrupt AI models.
The result is another turn in a long-running fight over art, AI training, copyright concerns, and control. For artists who use anti-AI protection tools, LightShed is a warning: a defense that works today may not work forever.
What LightShed Is Designed To Do
LightShed targets a layer of digital “poison” that artists can apply to online images. That poison is meant to make AI training systems misunderstand the work when they scrape images from the internet.
The researchers behind LightShed say they are not trying to steal artists’ work. Their concern is that artists may rely too heavily on protections that can later be removed without them knowing.
“You will not be sure if companies have methods to delete these poisons but will never tell you,” says Hanna Foerster, a PhD student at the University of Cambridge and the lead author of a paper on the work.
Foerster worked with researchers from the Technical University of Darmstadt and the University of Texas at San Antonio. The group will present its findings at the Usenix Security Symposium in August.
Why Artists Turned To Glaze And Nightshade
Generative AI image models need broad collections of visual material for training. The source article says those data sets allegedly include copyrighted art without permission, which has led artists to worry that models may learn their style, imitate their work, and threaten their livelihoods.
In 2023, researchers created tools including Glaze and Nightshade as potential defenses. Shawn Shan, who worked on these protections, was named MIT Technology Review’s Innovator of the Year last year for that work.
Glaze and Nightshade do not block an image from being seen online. Instead, they alter enough pixels to confuse an AI model while preserving the quality that a human viewer sees. These nearly invisible changes are called perturbations.
- Glaze is intended to make a model misunderstand style, such as reading a photorealistic painting as a cartoon.
- Nightshade is intended to make a model misunderstand the subject, such as reading a cat in a drawing as a dog.
- Glaze is used to protect an artist’s individual style.
- Nightshade is used to attack AI models that crawl the internet for art.
For artists, especially those with fewer resources, these tools can feel like a practical line of defense while regulation around AI training and copyright remains unsettled.
How LightShed Removes The Poison
LightShed was trained on pieces of art with and without Nightshade, Glaze, and similar programs applied. Foerster describes the process as teaching LightShed to reconstruct “just the poison on poisoned images.”
That matters because the tool is not trying to remake the whole artwork. It is trying to identify the added perturbation, determine how much of it is needed to confuse an AI system, and then remove that layer.
The source article describes LightShed as highly effective and more adaptable than simple methods other researchers have found. It can take what it learns from one anti-AI tool, such as Nightshade, and apply that knowledge to others including Mist or MetaCloak, even without seeing them ahead of time.
LightShed does have some trouble with small doses of poison. But the source article notes that smaller doses are also less likely to damage an AI model’s ability to understand the underlying art. That creates a difficult trade-off for artists using these protections.
Why This Matters For Digital Artists
Around 7.5 million people have downloaded Glaze. Many are artists with small and medium-size followings and fewer resources, making practical protection tools especially important to them.
LightShed does not mean artists have no options. It does suggest that existing anti-AI art protections should be seen as temporary technical barriers rather than final solutions.
The creators of Glaze and Nightshade appear to share that view. The website for Nightshade had already warned that the tool was not future-proof before work on LightShed began. Shan also argues that tools like his still matter because they create friction.
“It’s a deterrent,” says Shan.
As Shan frames it, the aim is to place enough obstacles in the way of unwanted use that AI companies find it easier to work with artists. He says “most artists kind of understand this is a temporary solution,” but still sees value in making unwanted scraping harder.
The Next Round Of Protection
Foerster hopes to use what she learned from LightShed to help create new defenses for artists. One idea mentioned in the source article is clever watermarks that remain with the artwork even after it has passed through an AI model.
She does not believe any such approach would protect a work against AI forever. But the goal is to shift some leverage back toward artists, even if only until the next round of the contest begins.
That is the central lesson of LightShed. The conflict over AI training and digital art is not only happening in courts, contracts, or culture. It is also happening inside the images themselves, where small technical changes can become tools of resistance, and where new tools can then try to erase them.