There's a contradiction in that list: proper code reviews require knowledge of the problem the code professes to solve. If AI is poor at context-dependent details, it can't shine at reviewing code.
It doesn't say the AI is poor at context dependent details. Specifically it says multi-file context or context dependent on knowledge of multiple domains. There's not necessarily any contradiction here.