Science News
When AI art has no author: Study finds generated images often can’t be traced to training data Read Full Article #AIscience #AIart #trainingdata
* 🤖 MIT researchers have discovered "attribution decay" in AI image generators, meaning that as models are trained on more data, the influence of any single training image diminishes to the point where removing it doesn't change the output.
* 🏗️ To overcome the prohibitive cost of retraining AI models, MIT CSAIL developed a "diffusion ensemble" architecture. This allows researchers to effectively remove the influence of specific training data by disabling components rather than re-engineering the entire model.
* ⚖️ This "attribution decay" has significant implications for copyright law, potentially meaning AI outputs are not derivative works and are themselves copyrightable, as they are not attributable to any single piece of training data.
#AI #Copyright #MachineLearning
(⚠️ Formatting disabled due to Markdown error)
* 🏗️ To overcome the prohibitive cost of retraining AI models, MIT CSAIL developed a "diffusion ensemble" architecture. This allows researchers to effectively remove the influence of specific training data by disabling components rather than re-engineering the entire model.
* ⚖️ This "attribution decay" has significant implications for copyright law, potentially meaning AI outputs are not derivative works and are themselves copyrightable, as they are not attributable to any single piece of training data.
#AI #Copyright #MachineLearning
(⚠️ Formatting disabled due to Markdown error)
🤔1