MIT researchers identify AI attribution decay complicating copyright infringement claims

The discovery of AI attribution decay introduces a significant technical hurdle for rights holders attempting to prove direct infringement in generative media. If a model can produce similar results without a specific copyrighted work in its training set, the legal link between source material and output weakens. This complicates the landscape for streaming platforms and studios developing proprietary generative tools, as it shifts the focus from output similarity to the legality of the initial training process itself. Industry observers should watch for how the U.S. Copyright Office integrates these findings into its case-by-case evaluations of human creative expression versus algorithmic generation.

The MIT study on attribution decay arrives amid a wave of high-profile copyright litigation targeting generative AI companies, several of which hinge on proving that specific copyrighted works directly influenced model outputs. In August 2025, a federal judge in the Northern District of California dismissed most of the copyright claims brought by artists against Stability AI and Midjourney, ruling that the plaintiffs had not adequately demonstrated substantial similarity between their works and the AI-generated outputs. That ruling underscored the evidentiary gap that attribution decay now quantifies: even if a work was present in training data, establishing a causal link to a specific output remains technically elusive. Meanwhile, the U.S. Copyright Office published the second part of its AI and digital replica report in January 2025, concluding that AI-generated outputs lacking sufficient human authorship are not copyrightable, a stance that further complicates the legal framework around generative media in streaming and entertainment workflows.

On the licensing and business front, major content companies are pursuing dual strategies of litigation and licensing deals with AI developers. Disney and Universal filed a joint lawsuit against Midjourney in June 2025, alleging that the platform's models were trained on thousands of copyrighted characters and scenes, marking the first time major Hollywood studios jointly targeted a generative AI company. The attribution decay finding could weaken the studios' ability to demonstrate direct copying, potentially shifting legal arguments toward the unauthorized use of copyrighted material in training datasets rather than output similarity. Separately, OpenAI announced a content licensing partnership with News Corp in May 2024, a deal valued at over $250 million annually, signaling that some rights holders are opting for commercial agreements rather than relying on infringement litigation. These parallel tracks suggest the industry is converging on licensing as the primary economic mechanism for AI training data, regardless of how attribution science evolves.

From a technical standpoint, the attribution decay phenomenon connects to broader research on machine unlearning, where the goal is to remove specific data points' influence from trained models. A 2025 study published in the Proceedings of the International Conference on Machine Learning found that current unlearning techniques fail to fully erase the influence of individual training samples in models with more than one billion parameters, corroborating the MIT team's findings at scale. For streaming platforms experimenting with for content creation, marketing assets, and personalized thumbnails, this technical reality means that compliance strategies based on dataset filtering alone may be insufficient. The research also intersects with ongoing efforts at the National Institute of Standards and Technology, where NIST released its AI Risk Management Framework Generative AI Profile in July 2024, providing guidance on provenance and transparency that could inform future regulatory approaches to AI-generated content in media distribution.