Biases contained within historical datasets makes addressing what underpins automated decisions difficult – defining what is ‘fair’ is key.