Generative AI has long been accused of copying artists’ work outright (see the numerous copyright lawsuits winding their way through court). But a new study out of MIT makes a very different argument—one that could complicate how those cases hold up.
In the study, published in Nature, the MIT researchers Zheng Dai and David Gifford set out to to test whether a generated image can be traced back to a single piece of training data. They were looking specifically at diffusion models, the systems most often used for generating images and video.
Their finding? It comes down to how big the training data set is. They found that the more data a model is trained on, the harder it becomes to attribute its output to any particular piece of training data. In fact, when they removed a specific piece of training data from the set, they found that it had little to no bearing on the updated output of the model. As the researchers put it: “We can often omit any sample or creator from the training data without affecting a generated sample.”
This trend, what the researchers call “attribution decay,” means that artificial intelligence can produce an image that resembles a particular artist’s work, while having no provable causal link to that artist’s actual contribution to the training data.
The researchers suspect that this happens because larger datasets tend to contain a lot of visual redundancy; many different images share overlapping features, so no single image is responsible for the DNA of an AI-generated image. They got the same result again and again, across dozens of experiments. But they’re careful to treat this as their best explanation rather than something they’ve directly proven (more on that below).
How this works
To get to this finding, the researchers needed a way to test cause and effect precisely, because diffusion models don’t work like a database you can just delete files from. During training, a model doesn’t store copies of the images it sees. Rather, it adjusts millions of internal numerical settings based on patterns across the entire training set, and later uses those settings to turn random noise into a new image. That means you can’t simply delete one training image and expect a straightforward before-and-after comparison. The influences of all the training images are normally tangled together across the whole model.
Instead, the researchers built models out of separate components, each trained independently on a different slice of the data, then combined. That structure let them cleanly remove the influence of one artist, person, or image by switching off just the components that had been exposed to it, without retraining the entire model from scratch. They call this technique “ablation.”
AI image generators learn by studying huge numbers of pictures and learning to reconstruct them. Dai and Gifford generated an image using the full training set, then regenerated it with one artist, person, or image removed, while everything else remained the same. If the newly generated results didn’t change, they could surmise that the missing piece wasn’t the cause of that particular output; the model would have generated a nearly identical image regardless.
They also tested the standard shortcut people use for spotting AI copying: Finding the training image that looks most similar to an output and calling it the source. That shortcut got it wrong more often as datasets grew; the “closest match” was frequently coincidental. There wasn’t causality because removing it often changed nothing in the generated results. That effect undercuts a lot of existing “AI copied this artist” claims, which rely on visual similarity rather than actually testing cause and effect.
Why does this happen? The researchers offer an explanation, though they’re careful to call it a conjecture rather than a proven fact. According to their paper, they “conjecture that attribution decay happens because features that are important to model behavior are distributively and redundantly encoded throughout its training set.”
The “unattributability” effect kicks in only at large scale. The paper states this effect is already significant at scales of 10,000 to 100,000 images, while many commercially deployed models train on datasets with up to a billion images.
The Warhol effect
To understand how this works in practice, let’s use Andy Warhol’s artwork as an example. Imagine a model trained on 50,000 works of art that happen to include Warhol’s silkscreens. If you delete Warhol’s specific images and regenerate, the output barely changes. This is not because the model “understands” Warhol’s style in some abstract sense, but because Warhol wasn’t the only artist doing bold flat colors, repeated grids, and pop culture subjects.
In fact, numerous other artists in that same dataset were doing visually similar things. That means that the visual features of thousands of images overlap enough that no one image is irreplaceable. In other words, Warhol’s paintings weren’t a unique ingredient; they were one of many sources that use the same visual pattern. Removing them would leave plenty of redundant signal behind for the model to draw from.
It’s important to note that the researchers are not arguing that AI models can simply create a Warhol-esque image out of thin air. If that same dataset had contained zero pop art, zero flat-color silkscreen work, and zero repeated-grid compositions, the model would have nothing to draw on to produce an image in that style.
Rather, they are saying unattributability kicks in when a style is redundantly present across many images. Delete one source of a common feature, and the feature survives through the remaining similar images. Delete every source of a rare feature, and the model loses the ability to produce it entirely.
What this means
These findings makes it more difficult to point to a specific artistic provenance, which could undermine artists’ arguments that AI copied their specific artwork. The same redundancy that lets a style survive one artist’s removal also makes it very hard to prove any single artist was the definitive source of a given output.
The peer-reviewed paper in Nature gives AI companies a sharp defense weapon in court, but not a shield. The researchers themselves tell about the legal stakes, with a careful hedge attached: Unattributability “ostensibly provides a refutation of access, a key element used in establishing infringement, thereby circumventing the intellectual property protections designed to limit such use as long as the harvest is conducted on a sufficient scale for attribution decay to manifest.”
That word “ostensibly” matters because even the authors are flagging this as a plausible legal argument, not a settled one. If the AI gives you a silkscreen of a brightly colored tomato can, it answers only one narrow question: Did ingesting Warhol’s art cause this image? It says nothing about whether scraping an artist’s work into a training set without permission was legal in the first place. That is the bigger fight already underway in dozens of lawsuits.
Overall, the research sharpens one argument in the AI copyright fight without definitely settling it. Companies now have real evidence against being blamed for specific outputs. But the bigger question, whether using copyrighted work to build these systems was ever allowed, remains exactly where it was before this paper.
