New research from China offers a non-invasive way to protect copyrighted animation characters, by ‘injecting’ generic substitutes at inference time. But can this method defeat the ingenuity of prompt-hackers? Given the extent of copyrighted material present in hyperscale datasets (because of the sheer cost of weeding such infractions out), models trained on them tend to be able to reproduce copyrighted characters. Direct editing of the trained model, or the use of filters to intercept keywords…