[ADITECH] Algorithmic Fairness in High-Stakes AI Systems (Sp)

I delivered a seminar co-organized by ADITECH and NAIR, focused on the technical and societal challenges of algorithmic decision-making systems. The session explored how machine learning models can introduce or amplify discrimination, particularly in high-stakes contexts such as hiring, credit allocation, or access to public services.

The talk combined:

• ⚖️ Algorithmic fairness foundations: definitions of fairness, impossibility results, and trade-offs.
• 🧮 Sources of bias: data imbalance, label bias, proxy variables, feedback loops, and distribution shift.
• 🛠️ Mitigation strategies: pre-, in-, and post-processing techniques.
• 🏛️ Regulatory and governance perspective: transparency obligations, risk categorization, and compliance considerations under the EU AI Act.
• 👥** Case study** (Viogén system): analysis of the Spanish VioGén algorithm used for gender-based violence risk assessment, examining transparency, risk scoring methodology, potential bias amplification, and governance challenges in public-sector AI deployment.

The discussion emphasized that fairness in AI is not only a technical optimization problem but a sociotechnical design challenge requiring alignment between modeling choices, institutional constraints, and legal frameworks.