Towards Fairness Certification in Artificial Intelligence

Discriminatory biases can enter AI systems through training data: The paper emphasizes that machine learning models are "data-driven" and multiple pathways exist for biases to "seep into data," which algorithms then risk amplifying rather than mitigating.

High-stakes decisions require formal fairness certification: In sensitive domains like healthcare, hiring, justice, and banking, the paper argues that AI systems need official guidelines and certification procedures before deployment—similar to how other critical technologies are regulated.

AI can amplify existing societal prejudices: The research warns that algorithms trained on biased data may "over associate protected attributes such as gender, ethnicity or disabilities" with predictions, effectively magnifying historical stereotypes rather than reducing them.

Fairness certification requires ongoing monitoring: Beyond initial approval, the paper proposes that certified AI systems need "conformity assessment procedures" to continuously monitor their functioning and ensure fair decisions over time, not just at deployment.

Establishing operational standards is urgent: Drawing on expertise in measurement standards and certification, the authors call for defining "operational steps needed for AI fairness certification," suggesting that formal frameworks for assessing and certifying AI fairness are currently lacking but essential.