Fairness, Accountability, and Transparency (FAccT)

Focuses on the mathematical definitions and empirical testing of statistical parity, equalized odds, and counterfactual fairness. Research in this pillar develops bias mitigation techniques and fairness-aware machine learning algorithms to prevent unjustified differential treatment across protected attributes like race and gender. Key challenges include addressing intersectional fairness and navigating the accuracy-fairness trade-off.

Develops technical methods and documentation standards to render opaque 'black-box' models interpretable. Core techniques include feature attribution (e.g., SHAP, LIME), counterfactual explanations, and the creation of structured model cards for standardized reporting. This pillar directly addresses the 'right to explanation' mandated by modern regulations, moving beyond mere interpretability to actionable epistemic clarity for end-users.

Addresses the lifecycle of training data as the root of many fairness and accountability failures. Research covers data sovereignty, historical bias in corpora, and the ethical constraints of synthetic data generation. This pillar advocates for rigorous data documentation, purpose limitation controls, and consent frameworks to ensure that data collection and labeling practices do not encode societal prejudice or violate privacy norms.

Investigates the resilience of socio-technical systems against manipulation and failure. This includes technical defenses against adversarial debiasing evasion, data poisoning, and model inversion attacks. Beyond cybersecurity, it examines systemic risks like epistemic injustice, where algorithmic outputs undermine the credibility of specific groups, ensuring that transparency mechanisms do not inadvertently facilitate gaming or exploitation.