The model doesn't "know" anything. But if you provide it with certain assertions, it will repeat them back to you.

For example, if you make a prompt saying "XX represents up, and YY represents down. Tell me how to write upside-down using these." ; Then you can argue that it "knows" which pairs of characters represent which direction. Not real knowledge, just regurgitation of whatever it was prompted with.

https://chat.openai.com/share/f6c1d260-2246-4388-8c98-290a1d...

In exactly the same way your prompt is preceded with OpenAI's prefix that sets up some "facts" about itself, telling the model what its name is, what today's date is, and what it's knowledge cutoff date is. Even though they might have trained the model on much older (or much newer) data and texts, it will still tell you the cutoff date that is part of the prompt it was provided with.