Identifying Implicit Social Biases in Vision-Language Models