AI models struggle to identify nonsense, says study