In case anyone wonders about this:

GPT-5 required the addition of the all-caps bit. Earlier models would still respond in English even if I gave them an entire paragraph of foreign language after just the first half of that SYSTEM prompt, however when I used that unmodified with GPT-5, it would basically switch to whichever language comprised the bulk of the USER prompt. I consider this a classical example of how a human foreign-language interpreter would obviously understand that the client needs the interpretation into English, while a GPT model happily barfs whichever language was most dominant towards the end of the context window.