This creates an infinite regress problem - bias detection systems are themselves trained on data reflecting human judgments about what constitutes "bias."