MIT Scientists Uncover“Attribution Decay” In AI Art

(MENAFN- USA Art News) New scientific findings exploring a phenomenon dubbed“attribution decay” are complicating the legal and ethical debates surrounding artificial intelligence (AI)-generated art. Researchers at the Massachusetts Institute of Technology's Computer Science and Artificial Intelligence Laboratory published their study in Nature Communications on August 18.

AI models are trained on billions of images, a process that has faced criticism and is the subject of several ongoing or impending lawsuits. Many perceive this training as a form of involuntary extraction, arguing that the resulting images violate numerous artists' copyrights.

However, the new research, conducted by scientists Zheng Dai and David K. Gifford, suggests a different perspective. They discovered that removing specific data from large training datasets did not alter the AI's output. This finding could support an argument that if a copyrighted image was part of the training data, but its removal had no impact on the generated image, then the AI-created image might not constitute copyright infringement.

Dai explained the methodology:“We developed a method for taking away one piece of the training sets and then regenerating the image as though that piece of training data didn't exist.” He elaborated,“And if you find that [the output] doesn't change much, then you can't attribute it to that piece of data, because it didn't have any influence on the final output.” In the context of large datasets for image generation, he noted that“there is no piece of data that you can take out that significantly alters the image, and that [leads] to this idea of unattributability.”

Hypothetically, even if David Hockney's painting A Bigger Splash ( 1967 ) or his entire body of work were removed from a dataset, an AI tool might still generate a similar image with the right prompts. Dai stated that such scenarios are speculative, but if the training data contained indirect references, derivative images, or echoes of A Bigger Splash, the AI could still produce a resembling image, like the famous painting at the Tate Britain. In such a scenario, where Hockney's original was not directly part of the dataset used for generation, a copyright infringement claim might be contested.

Dai also commented on a perceived irony, saying:“If you view [the] training off of people's data as [infringement], it is sort of ironic, the more infringement you do, the less infringing the output is.”

While AI tools can directly infringe copyright when tasked with replicating specific material, public sentiment against these models often overlooks the enormous scale of datasets involved. When algorithms create images from broader prompts, they draw from millions, even billions, of images. Dai acknowledged that their findings could be seen as a“get-out-of-jail-for-free card” for tech companies facing copyright infringement allegations. However, he added:“It certainly looks like it from one angle, but I think it is trickier than that.” He noted that if all artists collectively opted out of AI training, the models would cease to exist. Yet, in legal disputes, companies could remove a copyrighted image to demonstrate that its absence does not alter the output, using unattributability as a defense.

The study does not absolve the training process itself, meaning an AI company could still be held accountable for how its algorithms were trained, even if unattributability is argued for a specific output. Attribution decay adds another layer of complexity to the contentious legal landscape where public perception often holds that all AI-generated content violates copyright, and none of it can be copyrighted. Nonetheless, the United States Copyright Office maintains a case-by-case approach, assessing each AI-generated work for evidence of human creative expression to determine copyright eligibility.

Source: The Art Newspaper

MENAFN04092026005694012507ID1111622727