My mental model of how ChatGPT might superficially "understand" its own limitations would be because, for example, it might have been specifically trained by people who do understand the limitations of the system.