AI research identifies attribution decay in generative models

Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have identified a phenomenon called ‘attribution decay’ in generative AI models. The study suggests that as models are trained on increasingly large datasets, it becomes difficult to trace specific outputs back to individual pieces of training data. The researchers found that removing a single image, or even all images by a specific artist, often results in no measurable change to the model’s output, complicating efforts to assign responsibility or credit for AI-generated content.

In the creative industry, studios like Rotterdam-based The Phoney Club are navigating these technological shifts by implementing ethical guidelines for AI usage. The agency utilizes licensed AI twins of real models to place digital doppelgängers in fictional environments. To maintain a connection to human reality, the studio follows a policy of not retouching or altering the models’ natural imperfections, such as wrinkles or shadows, as a way to acknowledge the existence of the real person behind the digital likeness.