The article shows how to make machine learning systems fairer without using demographic information. By training models on related tasks with demographic data, fairness can be improved on new tasks even without specific demographic labels. This approach transfers fairness objectives across tasks and domains, enhancing fairness in various scenarios.