August 18, 2026, (Inside AI) — A new study from MIT CSAIL finds that for large generative image models, individual training examples often have no measurable influence on any specific output. The researchers call this phenomenon attribution decay, and it challenges the legal and technical basis for tracing AI-generated images back to their training data.
The work, published today in Nature Communications, shows that as training datasets grow, the counterfactual radius of a generated image shrinks along an inverse power law. In plain terms, removing any single image, every image by one artist, or every photo of one person frequently leaves the output unchanged.
This has direct implications for copyright lawsuits, licensing negotiations, and proposed regulations that assume a link between training data and generated content. The researchers argue that if removing data changes nothing, that data cannot be responsible for the output.
"If you take away a piece of data and the output of the model doesn't change, then that piece of data didn't affect the output," says Zheng Dai SM '21, PhD '24, former MIT CSAIL researcher and lead author. "So it doesn't make much sense to attribute the output to that piece of data. And if you then do this one at a time for every other piece of data and find that the output doesn't change for any of them either, then it doesn't make much sense to attribute the output to any one of them."
The team built a new architecture called a diffusion ensemble to answer a what-if question that previous methods could only approximate. Instead of retraining a monolithic model without each image, they trained many smaller components on different data slices. Switching off the components that saw a particular image creates a true counterfactual model, not an estimate.
"All previous methods were approximate," says MIT Professor David Gifford, an MIT CSAIL principal investigator. "They really could not absolutely show that deleting individual things did not change the output. This paper introduces the first method that is absolute. You're actually deleting the inputs and deleting all influences of the inputs. This is the first exact method for doing large-scale deletion efficiently and showing that the results don't change."
The researchers trained 24 ensembles on datasets ranging from 256 to over 160,000 images, drawn from seven public collections including CIFAR-10, CelebA, MetFaces, and ArtBench. The counterfactual radius consistently shrank as dataset size grew. The pattern held across pixel-level and semantic similarity metrics, with statistical significance.
Stress tests reinforced the finding. The team retrained 1,282 separate models at small scale using brute-force ablation and saw the same decay. They pinned the removed fraction in place, fixed epochs, used text-prompted and class-conditioned models, and tried four similarity metrics. The result survived every variation.
Attribution decay complicates copyright claims against AI generators
The legal stakes are substantial. Courts and regulators have been operating on the assumption that if a generated image resembles a copyrighted work, the resemblance reflects copying. This study suggests that for large diffusion models, similarity may often be coincidence.
"If attribution worked, it would reliably tell us whether similarities between a model's output and a copyright-protected work are due to copying or coincidence," says James Grimmelmann, a law professor at Cornell Law School and Cornell Tech. "But this paper provides reason to think that attribution will fail for interesting models. Instead, technologists and courts will need to resort to other methods for assessing copying."
Gifford frames the finding as bearing on whether model outputs are derivative works. He argues that these models are creative, not simply copying what they are fed. If outputs have no connection to any individual training example, that raises questions about fair use, copyrightability of the outputs themselves, and how authors get compensated.
"One way to think about this is that these models are creative. They are not simply copying what they are fed, but creating brand new outputs. If those outputs have nothing to do with any individual piece of training data, that raises questions about fair use, about whether the outputs are themselves copyrightable as novel works, and about how authors get compensated when what comes out of a model isn't attributable to anything on the internet."
He also frames the ability to produce guaranteed unattributable outputs as an obligation for industry, not a loophole. Companies claiming their outputs are not derivative of the internet should revise their models to demonstrate this property, he argues.
"In order for these companies to claim their outputs aren't derivative of the internet in a copyright-infringing way, they need to revise their models to take advantage of the advances in this work, so they can show they're not creating derivatives of individual people or items."
The study focuses on diffusion models, which now dominate audiovisual generation and are used in scientific applications like protein structure modeling and therapeutic discovery. Whether the same attribution decay holds for large language models remains an open question, one that matters for the highest-profile copyright litigation in AI.
One practical finding emerged from the ensemble approach. The more training data, the better the ensembles held up against conventional single-model diffusion models. This suggests ensembles may be more data-efficient at scale, even beyond their use for attribution analysis.
"When you have low amounts of data, they do very poorly," says Dai. "But if you have more data, it actually scales better compared to the vanilla diffusion model."
The work was supported by Schmidt Futures. The open-access paper is available in Nature Communications.
