AI companies often assure musicians that their work can be removed from a model after training, but a new research initiative argues that such promises are technically hollow. Chris Castle, founder of the MusicTechPolicy blog, has launched the Machine Unlearning Research Hub to document the gap between industry claims and the current state of computer science.
Training a neural network on a song alters trillions of internal mathematical parameters across the system. The process has been compared to trying to extract eggs from a fully baked cake: once data is ingested, it cannot be simply deleted without retraining the entire model from scratch.
“Creators should not assume that a contractual promise to ‘remove’ works from future training is equivalent to proving that an existing model has actually forgotten them,” Castle said.
The Hub serves as an open knowledge resource tracking peer-reviewed computer science, audio research, and policy analyses on machine unlearning, the theoretical process of making AI models forget specific training data without a full rebuild.
Castle advises musicians not to trade upfront permission or payment for a promise of later deletion. Until true machine unlearning becomes technically feasible, he argues, obtaining strict pre-training consent remains the only reliable protection.
Castle said the initiative is meant to prevent musicians from being sold a “fake delete button” by grounding artist rights in verifiable research.
