Spotting bias in machine learning