Gemini doesn't know anything. All of its outputs are synthesized via pattern matching of the prompt against its training data. No one knows exactly what the sources of any given LLM synthesis are. If one asks for a summary of a specific article then it will do that, but that wasn't the prompt.