If so, maybe that says something about that claim?

I think there's probably a big difference between how data is stored in art models versus how LLMs store textual information.