Yes, but all that does is locate "I don't know" near the cooked up data within the embeddings. This doesn't actually reflect an absence of data in the training.