Protecting artists against ai: Withdrawal and Machine Unlearning

With certain methods, data can be removed and forgotten by a machine, called machine unlearning. Largely, this has been theorized for the purpose of correcting models which are biased or for security. In this context, however, it can be useful for removing large portions of copyrighted artwork, through comparisons of AI-generated images and copyrighted non-AI-generated art. This can be done through several methods, such as training non-copyrighted works of a similar nature or removing the influence of copyrighted works on the model. Despite this, the main goal of unlearning is to retain as much model usability as possible, which would still mimic and regurgitate artists’ styles. The same issue of contacting all artists whose works are in a model’s dataset would also be relevant here, as creative works created recently are technically protected by copyright by default (though this is not the same as registering ownership). To continue, it has been shown that despite a concept being erased from a model, it’s still possible to generate something directly related to that concept. Machine unlearning technology is still incredibly finicky as of now, with many people receiving mixed results.