"The relative merits of ‘poisoning’ techniques to protect user-made artwork from generative models. OpenAI, Microsoft and smaller ‘AI’ companies have drawn increased scrutiny for their silent seizure of artists’ creative property, necessary to sustain the past year’s influx of model-generated images/slop. Glaze/Nightshade – software that ‘poisons’ images, thereby preventing them from being used by models – have thus been hailed as a panacea. Yet the dubious efficacy and morality of both products may ultimately do further harm. I first will provide an exposition of the mechanism girding Glaze/Nightshade. Both rely on ‘adversarial perturbations’ – the addition of subtle quantities of noise into an image that causes misclassification. I then will discuss their historical evolution, increasingly widespread adoption and their beneficial consequences (offering autonomy for artists, forcing ‘AI’ companies to negotiate terms with artists). I will then discuss their inconspicuous downsides. Efficacy: Glaze has already been made redundant by some model architecture, and it is reasonable to expect that Nightshade will be similarly short-lived. Therefore these tools can only be accepted as temporary solutions, and require continuous development. Damaging useful models: Certain models depend on user-generated data for ostensibly benevolent aims: for example, models for medical classification which are trained on images of afflicted people. ‘Poisoning’ techniques threaten to disrupt said models, whose usage may play a crucial role in diagnostics. I will then connect this particular issue to the broader case of American copyright law and the need for some degree of federal intervention."
------Daniel Kyte-Zable
