That’s not necessarily true: arstechnica.com/…/googles-bard-ai-can-now-write-a…. If the question gets interpreted correctly and it manages to write working code to answer it, it could correctly answer questions that it has never seen before.