They can read context and with fairly high accuracy say whether that context contains enough information to answer a posed question. They cannot (and we cannot for them) introspect their own weights to say whether their weights already encode information sufficient to answer a posed question.