This article seems to be reading a lot into the word choice of a spokesperson, who is almost certainly very far removed from the model development process
I consider it pointless. All you need to do is train new models to ignore the introduced distortion. Learning how to ignore the distortion will make the models more robust.
Then a new distortion will be created that avoids the distortion detection of the new models. Ad infinitum.
I think it’d be nicer if we could just mark web images as ‘do not train on’?
The point is for this to be used as a defense against scrapers that would ignore a flag like that. I think if you asked the nightshade devs they'd say we should have both.