Public administrations deploy AI for welfare decisions affecting millions. Existing fairness definitions are static and miss the feedback loop between human reviewers and model retraining. We present ARIA and TFAS – the first fairness definition incorporating bounded human-reviewer feedback as a formal parameter. We prove the Temporal Injustice Theorem: if feedback coefficient δ≥1, demographic disparity diverges monotonically regardless of local correctness. Verified on n=100,000 synthetic applications with 30/30 unit tests. Implements EU AI Act Article 14.
Building similarity graph...
Analyzing shared references across papers
Loading...
Spilios Dimakopoulos
Building similarity graph...
Analyzing shared references across papers
Loading...
Spilios Dimakopoulos (Sat,) studied this question.
www.synapsesocial.com/papers/69eefd9bfede9185760d444b — DOI: https://doi.org/10.5281/zenodo.19764114