Even the best AI models studied can be fooled by nonsense sentences, showing that “their computations are missing something about the way humans process language.”