Large language models pass three-party Turing test, outscoring actual humans