openai claims o1 ai model is 'virtually perfect' at bias correction – but does the data agree?