When an artificial intelligence image generator produces a portrait, the question of whose work contributed to that output sits at the center of lawsuits, licensing negotiations, and proposed regulations worldwide. Artists demand credit and compensation. Companies seek legal clarity for their multibillion-dollar investments. Policymakers across jurisdictions are scrambling to assign responsibility for content generated by systems trained on vast swaths of the internet.<\/p>\n
New research from the Massachusetts Institute of Technology’s Computer Science and Artificial Intelligence Laboratory (CSAIL) suggests that for models trained on large datasets, this question may often have no meaningful answer. The issue is not merely that existing tools for tracing influence are inadequate. According to the study, the causal connection between individual training examples and specific outputs has, at sufficient scale, effectively disappeared.<\/p>\n
The MIT team has identified a phenomenon they term “attribution decay.” The more data a generative model is trained on, the less any individual training example matters to any particular output. This finding runs counter to intuitive assumptions about how AI systems learn and create. At sufficiently large scales, the researchers demonstrate, removing any single image from the training data \u2014 or every image by a given artist, or every photograph of a given person \u2014 produces no measurable change in the generated sample.<\/p>\n
If removing something changes nothing, the researchers argue, that piece of data cannot reasonably be said to be responsible for anything the model produces. This has profound implications for copyright litigation, fair use doctrine, and the emerging regulatory frameworks for generative AI.<\/p>\n
“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 on the work. “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.”<\/p>\n
Previous approaches to attribution have relied on approximations. They estimated a training example’s influence rather than actually removing it and retraining the model from scratch. The problem has always been computational: to truly test what a model would have produced without seeing a particular image, one must retrain the entire model without that image. Doing this for every image in a dataset of millions is mathematically and practically prohibitive.<\/p>\n
“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.”<\/p>\n
Dai and Gifford’s research, published in an open-access paper in Nature Communications<\/i>, details an architectural solution they built specifically to answer this question: a “diffusion ensemble.” Rather than training one monolithic model on all available data, this architecture consists of many smaller components, each trained on a different slice of the dataset. To determine what the model would produce without a particular image, one simply switches off the components that saw that image. No retraining is needed. No approximation is employed. The result is a true counterfactual model \u2014 not an estimate of one.<\/p>\nHow the Diffusion Ensemble Architecture Works<\/h3>\n
The team needed to verify that their clever architecture still functioned effectively as a generator. They put their ensembles head-to-head with 24 conventional diffusion models trained on the exact same data. By standard quality measures, the images produced by the ensembles were comparable. One finding surprised even the researchers: the more training data used, the better the ensembles performed relative to their single-model counterparts. This suggests the diffusion ensemble approach may actually be more data-efficient at scale.<\/p>\n
“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.”<\/p>\n
With the ablation method working reliably, the researchers could finally ask their central question at meaningful scale. They took a single generated image and imagined every alternate version of it that would have been produced by removing a different piece of the training data. The team calls this collection the image’s “counterfactual universe.” The distance between the original image and its most different alternate \u2014 termed the counterfactual radius \u2014 captures the maximum extent to which any single piece of training data could have mattered.<\/p>\n
The researchers trained 24 ensembles on datasets ranging from 256 images to more than 160,000, drawn from seven public collections including CIFAR-10, CelebA, MetFaces, and ArtBench. The pattern was consistent across all datasets: the larger the training set, the smaller the counterfactual radius. The relationship followed an inverse power law. This held whether differences were measured pixel by pixel or by semantic meaning, with statistical significance in both approaches.<\/p>\n
The team rigorously tested their own result against alternative explanations. Perhaps the ablation architecture itself was causing the effect? They redid the experiment the brute-force way at smaller scale, training 1,282 separate models from scratch. The attribution decay appeared regardless. Maybe larger datasets simply make each removal proportionally smaller? They pinned the removed fraction constant across dataset sizes, and the decay persisted. They tested fixed epochs, text-prompted models, class-conditioned models, and four different similarity metrics. The finding survived every variation.<\/p>\n
The implications of attribution decay run in directions that surprised even the researchers themselves. Gifford sees the finding as bearing directly on the legal question of whether model outputs constitute derivative works \u2014 one of the central issues in pending copyright litigation against companies like OpenAI, Stability AI, and others.<\/p>\n
“One way to think about this is that these models are creative,” Gifford says. “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.”<\/p>\n
The research also demonstrates how to produce outputs that are guaranteed to be unattributable to any specific training example. Gifford frames this capability as an obligation for the industry rather than a loophole to exploit.<\/p>\n
“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.”<\/p>\n
The current research specifically examines diffusion models, which have become the dominant architecture for generating audiovisual media. These models are also increasingly prevalent in scientific applications, including protein structure modeling and therapeutic discovery. Whether the same attribution decay holds for large language models \u2014 the technology at the center of the highest-profile copyright litigation involving companies like Microsoft, OpenAI, Meta, and Google \u2014 remains an open question requiring further investigation.<\/p>\n
James Grimmelmann, a law professor at Cornell Law School and Cornell Tech, offers perspective on the legal implications: “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. 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.”<\/p>\n
The timing of this research is significant. Multiple class-action lawsuits from artists, authors, and photographers allege that generative AI companies infringed copyright by training on their work without permission or compensation. Several major publishers, including The New York Times, have filed suit. Getty Images has pursued litigation against Stability AI over the use of its copyrighted photographs in training data. Meanwhile, companies have begun striking licensing deals \u2014 Shutterstock with OpenAI, Reddit with Google, and numerous publishers with various AI firms \u2014 to establish authorized training data pipelines.<\/p>\n
The MIT finding suggests that these legal and commercial frameworks may rest on a questionable assumption: that individual training examples have a meaningful, traceable influence on specific outputs. If attribution decay is a general property of large-scale generative models, then the entire enterprise of assigning credit or blame to specific training data becomes technically problematic.<\/p>\n
This does not necessarily mean that training on copyrighted data is legally permissible. The question of fair use involves multiple factors beyond technical attribution, including the purpose of the use, the nature of the copyrighted work, the amount used, and the effect on the potential market. But the MIT research undermines one common argument made by both sides: that specific similarities between a model’s output and a training example can be attributed to copying rather than coincidence or lawful transformation.<\/p>\n
There is a counterintuitive privacy dimension to attribution decay. On one hand, the finding suggests that individuals whose images or text appear in training data cannot point to specific outputs and claim those outputs are “theirs” in any technically meaningful sense. This could strengthen the legal position of AI companies defending against claims of derivative copying.<\/p>\n
On the other hand, the same research shows that at smaller scales \u2014 for models trained on niche datasets, or for fine-tuned models adapted to specific domains \u2014 attribution may still be possible. This creates a regulatory paradox: the largest, most widely used models may be the least attributable, while smaller specialized models might still have traceable training influences. Policymakers will need to grapple with whether and how to regulate based on model scale and training data size.<\/p>\n
The research was supported by Schmidt Futures. The team’s methodological innovation \u2014 the diffusion ensemble architecture \u2014 allows for exact counterfactual analysis that was previously impossible at scale. By training many smaller component models on disjoint subsets of data, the researchers can simulate removing any training example without the computational cost of full retraining. This enables precise measurement of how much each training example actually contributes to each output.<\/p>\n
The team’s finding that the counterfactual radius shrinks according to an inverse power law as dataset size grows has significant predictive value. It suggests that as AI companies continue to scale their models and training data \u2014 an industry trend that shows no signs of slowing \u2014 the attribution problem will only become more acute. Current state-of-the-art models like DALL-E 3, Midjourney, and Stable Diffusion are trained on datasets containing billions of images. At these scales, the MIT research implies, no single image is likely to have any detectable influence on any particular output.<\/p>\n
The MIT study forces a reckoning with fundamental questions about generative AI. If these systems are genuinely creative in the sense of producing outputs that are not attributable to any specific input, then the legal frameworks built on copyright and derivative works may be poorly suited to regulate them. The models may be doing something that copyright law was not designed to address: learning patterns and styles from vast quantities of data and then generating novel combinations that cannot be traced back to any source.<\/p>\n
This does not resolve the question of whether training on copyrighted data without permission or compensation is ethical or legal. Artists whose work is used to train commercial models may still have legitimate grievances, even if no specific output can be attributed to their specific images. The industry may need to develop new compensation models \u2014 perhaps based on aggregate data contributions rather than individual attributions \u2014 that fairly reward creators for their role in training these systems.<\/p>\n
Gifford’s framing of the industry’s obligation is pointed: companies that want to claim their models do not produce derivative works must adopt architectures that can prove it. The diffusion ensemble approach provides one path to making that guarantee. Whether companies will voluntarily adopt such methods, or whether courts and regulators will require them to do so, remains to be seen.<\/p>\n
The research also opens questions about how to think about AI creativity. If attribution decay means that large generative models are genuinely synthesizing new patterns rather than recombining memorized fragments, then the outputs might be entitled to copyright protection as original works \u2014 a possibility that current U.S. Copyright Office policy largely rejects. The technical reality that the MIT team has uncovered may eventually force a reshaping of legal doctrine.<\/p>\n
Dai and Gifford’s work, while focused on diffusion models and image generation, points toward a broader principle: at scale, generative AI systems may be more than the sum of their training data in a way that makes individual examples truly irrelevant to any specific output. For the artists demanding credit, the companies seeking clarity, and the policymakers trying to assign responsibility, this finding suggests that the question they are asking \u2014 whose work went into this particular AI-generated image? \u2014 may simply not have an answer.<\/p>\n","protected":false},"excerpt":{"rendered":"
When an artificial intelligence image generator produces a portrait, the question of whose work contributed to that output sits at the center of lawsuits, licensing negotiations, and proposed regulations worldwide. Artists demand credit and compensation. Companies seek legal clarity for their multibillion-dollar investments. Policymakers across jurisdictions are scrambling to assign responsibility for content generated by […]<\/p>\n","protected":false},"author":7,"featured_media":76913,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"fifu_image_url":"https:\/\/pub-4d4fc17555de4152be07eaf2a416a31e.r2.dev\/en\/ocie_1787107078177.jpg","fifu_image_alt":"MIT Study Reveals AI Art Lacks Traceable Training Data","footnotes":""},"categories":[31],"tags":[],"class_list":["post-76912","post","type-post","status-publish","format-standard","has-post-thumbnail","category-technology"],"fifu_image_url":"https:\/\/pub-4d4fc17555de4152be07eaf2a416a31e.r2.dev\/en\/ocie_1787107078177.jpg","fifu_image_alt":"MIT Study Reveals AI Art Lacks Traceable Training Data","_links":{"self":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/76912","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/users\/7"}],"replies":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/comments?post=76912"}],"version-history":[{"count":0,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/76912\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media\/76913"}],"wp:attachment":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media?parent=76912"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/categories?post=76912"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/tags?post=76912"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}