How to prevent human bias from infecting AI