Hidden Backdoors in Human-Centric Language Models