Automatic evaluation methods based on text overlap and language model judgments can approximate human ratings under specific conditions, but their reliability is highly context-dependent.
