The data used to train AIs reflect societal stereotypes and prejudices, for example, towards underrepresented groups. In order to maintain the confidentiality of sensitive data, such as health data, while ensuring that it is not biased, learning methods must be adapted.