Exploring the Limitations of Detecting Machine-Generated Text