This. The models are out there. Maybe they will just be illicitly shared but even if no new models are trained from scratch I suspect there will be many ways to use extend existing models without going back to scraped images.

I always felt that we already had a solution - I can already get all those images from a web search. Where the law currently intervenes is when I try and distribute works based on close copies of them. Why is this insufficient?

Sure. And neither are mp3s of the same songs that were blowing up on Napster.

The existence of widespread illegal means to procure something doesn't mean that we don't and shouldn't require legitimate businesses to abide by the law or require them to make amends for their current transgressions.