Its updated training data has numerous articles about Elon promising to “fix” Grok after giving a response that he didn’t agree with, so to avoid that, it’s checking Elon’s tweets when answering questions related to those. That’s my theory at least.