MIT CSAIL Finds Scaling AI Models Erases Image Attribution

Scientists at the MIT Computer Science & Artificial Intelligence Laboratory (CSAIL) have identified a phenomenon they call attribution decay in large-scale generative diffusion models. Lead author Zheng Dai and his team found that as these models grow and ingest more training data, the influence of individual inputs vanishes. To prove this, the researchers used a novel ablation method on a diffusion ensemble architecture, allowing them to swap out data components without retraining the entire system.

The team trained 24 ensembles on datasets ranging from 256 to more than 160,000 images. They pulled these samples from seven public datasets, including ArtBench, CIFAR-10, Fashion-MNIST, CelebA, and MetFaces. In one test, a model trained on public domain art from 744 artists generated virtually identical images even after specific artists were entirely removed from the training set. This demonstrated that the system could recreate styles and images it had never actually seen during training.

This finding directly impacts ongoing legal battles, such as the California class action lawsuit against Stability AI and Midjourney, and Getty Images' litigation against Stability AI in the UK. If a model's output cannot be traced back to a specific copyrighted input, proving infringement becomes much harder. Co-author David Gifford noted that this suggests models are generating entirely novel works rather than replicating training data, which could redefine fair use and copyright eligibility for AI-generated works.

For AI developers and practitioners, this shift means the industry must focus on proving their models do not generate derivative works. Gifford suggested that building models guaranteed to produce unattributable outputs is an obligation for the industry. Developers will need to design architectures that actively prevent the replication of individual copyrighted works, shifting the compliance focus from data-cleansing to architectural verification.