It's because the model doesn't actually see text during training or prompting, it gets a compressed form called BPE instead, and the compression is fixed rather than learned.