Policy makers are starting to understand that many systems running on AI exhibit some form of racial bias. So they are happy when computer scientists tell them that ‘debiasing’ is a solution for these problems: testing the system for racial and other forms of bias, and making adjustments until these no longer show up in the results.
European Digital Rights (EDRi) commissioned a report, asking academic researchers whether ‘debiasing’ is indeed a feasible way towards a more equitable use of AI. Turns out it isn’t.
In the report, Agathe Balayn and Seda Gürses point out the limitations of ‘debiasing’. Their main concern is that a focus on ‘debiasing’ shifts political problems (of structural discrimination) into a technical domain, which is dominated by the large commercial technology companies.
- The machine learning view – Some aspects of machine learning are inherently harmful.
- The production view – the making of AI systems have potential harmful effects that fall outside of the system.
- The infrastructural view – the computational infrastructure that is needed for AI systems is in the hands of few, creating power imbalances.
- The organizational view (AI will automate and centralise workflows affecting the structure of the public sector and democracy.
See: If AI is the problem, is debiasing the solution? at European Digital Rights (EDRi).