We introduce modal aphasia, a surprising and systematic dissociation in which unified multimodal models demonstrate strong capabilities for generating visual content while simultaneously failing to access that same knowledge through text queries.

We measure the frequency at which LLMs reproduce training data when not prompted to do so adversarially, and find that it can happen frequently even on accident.

We find that empirical evaluations of heuristic privacy defenses can be highly misleading, and propose a new evaluation protocol that is reliable and efficient.