SAPS sounds alarm over AI-generated police images

A new study from the Massachusetts Institute of Technology’s Computer Science and Artificial Intelligence Laboratory has uncovered a phenomenon known as "attribution decay." The researchers, Zheng Dai and David K. Gifford, discovered that in large datasets, removing certain data did not alter the output generated by AI models. This finding complicates the ongoing debate over the copyright infringement of AI-generated images, as it suggests that even if a copyrighted image was part of the training data and subsequently removed, the AI model could still produce a similar image based on other data it has encountered.

The researchers developed a method to remove one piece of training data and regenerate the image, demonstrating that the output is largely unaffected. This "unattributability" raises questions about the validity of copyright infringement claims against AI models, as the training data may contain traces, derivatives, or echoes of the original copyrighted work, making it difficult to attribute the infringement to a specific piece of data.

While the study may provide a potential loophole for tech companies facing copyright infringement allegations, it is essential to note that AI models still rely on a vast array of data sources, including public domain works and derivative creations. The legal landscape surrounding AI-generated art remains contentious, with copyright offices emphasizing the need for a case-by-case analysis of human creative expression in the generated works.