What is a fair model?

How can fairness be concretized when assessing a model?

Fairness (according to the Dutch DPA) must be interpreted according to the circumstances of a case and existing standards. The data controller must be honest about how personal data is used and must make sure that no undue negative effects are created for the data subjects. For this principle, the entire process must be fair to all data subjects; one part of the process cannot offset the degree of fairness of another part of the process. When assessing the outcomes of a model, an important factor is whether those involved perceive the outcomes to be fair. Fairness thus goes beyond lawfulness in the context of the GDPR and must be justified and accounted for by the data controller.

See also: https://autoriteitpersoonsgegevens.nl/sites/default/files/atoms/files/toezicht_op_ai_en_algoritmes.pdf