Indeed, how can anyone? Maybe one could build some models to counter for internal biases, but at this point that's probably the very best one could hope for.