OpenAI’s VP of global affairs claims o1 is ‘virtually perfect’ at correcting bias, but the data doesn’t quite back that up