You fed it a handful of "I'll never forgive you, liar" prompts beforehand. That's the type of language used by a couple having a fight. All it did was respond accordingly.

Except there's no way to prove it provided a real document and didn't just make it all up (i.e. doing what an LM is supposed to do).