Language models have the same defect - they are quite brittle and susceptible to black-box adversarial attacks ( eg arxiv.org/abs/1907.11932 )