The answer is, as is often the case, "yes".

In some cases, an LLM may truly "consider the architecture" internally, within its latent representations, and in others, it can output a similar phrase simply because it's "expected" of it.

"Where" is pretty clear. There aren't that many places within an LLM, and hidden state is the main culprit. How to read that space is another matter entirely.