MIT researchers identify AI attribution decay complicating copyright infringement claims

The discovery of attribution decay fundamentally challenges the legal framework for intellectual property in generative media. If diffusion models can recreate proprietary styles without direct access to the source, the argument that AI outputs are derivative works becomes harder to prove in court. This shift favors AI developers like Stability AI by framing their models as creative rather than extractive, potentially shielding them from licensing demands. For the streaming and digital media ecosystem, this complicates the development of automated rights-clearing tools and audit trails for synthetic content. Watch for how federal courts in California adjust evidentiary standards for 'machine unlearning' in upcoming artist-led class action suits.

Stability AI and Midjourney face mounting legal pressure as courts grapple with the technical realities of how diffusion models learn and reproduce visual content. In January 2025, Getty Images and Stability AI reached a settlement that included a licensing agreement for Stability's commercial products, resolving one of the highest-profile copyright disputes in generative AI. The deal granted Stability AI a license to use Getty's library for training future models, establishing a precedent that major rights holders may prefer licensing revenue over outright bans on training data usage. That settlement came after a UK court ruled in November 2024 that Getty's US copyright claims could proceed, narrowing the case but keeping the core infringement question alive.

The regulatory landscape around AI training data is tightening across multiple jurisdictions. The EU AI Act, which entered into force in August 2024, requires providers of general-purpose AI models to publish detailed summaries of training data content and to comply with copyright opt-out mechanisms under the Digital Single Market Directive. The European Commission published guidelines in July 2025 clarifying that AI developers must demonstrate they respected machine-readable rights reservations from rights holders, a requirement that directly intersects with the attribution decay problem identified by MIT researchers. If models cannot reliably trace outputs to specific inputs, compliance verification becomes significantly more complex for both regulators and developers. In the United States, the Copyright Office released a report in January 2025 stating that AI-generated outputs lacking sufficient human authorship cannot receive copyright protection, while leaving open the question of whether training on copyrighted works constitutes fair use.

Technical approaches to proving or disproving provenance in diffusion model outputs are advancing alongside the legal debates. Researchers at the University of Maryland demonstrated in a 2025 paper that membership inference attacks, which attempt to determine whether a specific image was in a model's training set, achieve accuracy rates below 60% on models with more than 100 million parameters, reinforcing the MIT CSAIL finding that scale erodes attribution signals. Meanwhile, the C2PA specification version 2.4 introduced a new JSON-based serialization format called crJSON alongside expanded asset format support and new assertions for live video, creating a parallel provenance layer that does not depend on model-level attribution. For streaming platforms evaluating synthetic content policies, the convergence of attribution decay research and C2PA deployment suggests that watermarking and metadata standards, rather than model interrogation, will become the primary mechanism for rights verification in AI-generated video.