Identifying Bias in Machine-generated Text Detection