I believe this was a joke. Papers compare trained models to "ground truth" to determine accuracy. Ground truth refers simply to the correct answers, but, the way it's used grammatically, you can interpret it as another model called "ground truth" which has a 100% accuracy at all tasks.