protecting artists against ai: Watermarking and MEtadata

There are two types of watermarking, visible and invisible watermarks. Visible watermarking refers to a logo or image superimposed on top, while invisible watermarks are embedded within the host data and have no effect on the visual characteristics of the image. There are many applications for digital watermarking, however, they can be attacked using simple image modifications, such as rotation, cropping, and other minor changes. Therefore, digital watermarks need to be built in a way that maximizes robustness, that is, they must be able to withstand all sorts of edits. Because of this, watermarks, even ones within an image’s host data, are difficult to fully rely on.
    In a similar vein, a string or tag in an image’s metadata could also be used to signify whether or not the owner indicates consent for training in the future. As it is relatively simple to modify an image’s metadata using text, this method is accessible towards artists as well. The issue lies in the plethora of untagged images. It is highly likely that most datasets will continue to use images that neither prohibit nor allow use in AI models, as not using them would mean a drop in model efficiency. With the many inactive artists whose images are still on the internet and used for training, it is not certain how effective this method will be either. Though metadata tagging seems promising, lack of knowledge or exposure surrounding it may also prevent it from taking off or becoming widely prevalent.