It's very hard to know because we don't know and can't experiment with its training data. So - it may be doing first principle reasoning, or it may be doing token substitution vs. some known example that it's seen before and is matching to.