Then simply train the model to predict whether a human can interpret it as a unicorn.