Artists Want to Prove Their Work Was Stolen by AI. A New Study Says That’s Impossible

A new study published in Nature Communications found that it is impossible to draw a direct causal link between a specific artwork and the output of diffusion-based image generators like Midjourney and Stable Diffusion, a phenomenon the researchers call 'attribution decay.' The study, led by MIT CSAIL PhD student Zheng Dai and MIT computer scientist David Gifford, trained two dozen custom models and found that as training set size increases, generated images become completely 'unattributable' to any particular training image, face, or artist. This poses a significant challenge for artists seeking to prove their work was used without permission in AI-generated images.

One way to figure out how something works is to take it apart one piece at a time, and see how it changes along the way. That was the approach taken by two AI researchers who wanted to understand how image-generating AI models arrive at their final outputs: Do they reference a particular image, the way a human portrait artist might look at a human subject while painting? Or is there some other process at play? Their results were surprising—and they could have big implications for artists trying to sue AI companies on grounds of intellectual property theft. Published last week in the journal Nature Communications , the study https://doi.org/10.1038/s41467-026-75667-5 found that in the case of diffusion models—the AI systems powering image-generating tools like Midjourney and Stable Diffusion—it’s impossible to draw a direct, causal link between a particular work of art, human face, or even body of work from a particular human artist, and the images created by the model. The sheer scale of the training dataset obscures its creative process; it’s like an astronaut aboard the International Space Station trying to spot a single tree in the Amazon Rainforest with her naked eye. The models’ outputs are therefore completely “unattributable,” according to the study’s authors: Even if they look exactly like the work of a human artist, you can remove that work from the training data and the output will remain the same. ‘Attribution decay’ Attributability is crucial for understanding both how diffusion models work, and for determining whether a human artist’s work was used to generate an image. “If a diffusion model generates something, you want to be able to say, ‘Oh, this part of the training data was responsible,” Zheng Dai, a fourth-year PhD student at the Massachusetts Institute of Technology’s Computer Science and Artificial Intelligence Laboratory CSAIL , and the study’s lead researcher told Gizmodo. To that end, Dai and his coauthor, MIT computer scientist David Gifford, set out to see how AI-generated images would be altered if you remove particular images, or segments of images, one at a time from their training datasets. The idea was that by quantifying the difference between the original AI-generated image and the ones created after the training images had been removed—which they call the “counterfactual radius”—you can determine the degree to which a particular part of the training dataset influenced the original. “In this alternative universe where that data didn’t exist, then your sample should be different,” says Dai. Rather than using publicly available models, whose training datasets aren’t publicly available, Dai and Gifford built two dozen custom models trained on anywhere from a few hundred to hundreds of thousands of images, all of which were taken from online image databases. The researchers then ran a series of experiments to find out how the models’ outputs would differ if they went in and removed particular images. They also ran the same experiment on models trained exclusively on people’s faces, and again on others trained on the work of different artists. In all three cases, the result was the same: The more images a diffusion model is trained on, the more difficult it was to determine which of those images went into the AI-generated output—a phenomenon they call “attribution decay.” The study “demonstrates that at large training set sizes it not only becomes infeasible to attribute generated images to training images, it also becomes infeasible to attribute generated people to the real people the model was trained on, and to attribute generated artwork to the artists the model was trained on.” Alien intelligences This is bad news for artists trying to sue AI companies: No matter how much an AI-generated image might look like something they themselves created, the new study suggests it’s impossible to know with complete certainty which images were referenced by the model to create its finished output. As Zheng and Gifford note in their report, commercial diffusion models are many orders of magnitude larger, and thus far less attributable, than the test models used in the study. For larger models, AI-generated images do not have any single human-generated inspiration. It’s a reminder of just how alien these black box systems are, even though the neural networks they’re based upon were designed to mimic processes in our own brains. When humans create something, whether it’s a painting or a novel, we’re usually conscious of the other people and works of art that are influencing us. We may even look directly at some kind of external reference, as in the case of the portrait artist. Diffusion models, in contrast, use the totality of their training data to generate new images in ways that are still in many ways deeply mysterious to us. The study also underscores just how far legal experts still have to go to untangle the knotty intellectual property and accountability questions these systems raise. But just because a task is difficult doesn’t mean it’s worth giving up on. “It’s important for us to understand how these models work to properly study them or regulate them,” says Dai.