The A lgorithmic F airness through the Lens of T ime (AF T ) workshop aims to spark discussions on how a long-term perspective can help build more trustworthy algorithms in the era of expressive generative models . Fairness has been predominantly studied under the static regime, assuming an unchanging data generation process. However, these approaches neglect the dynamic interplay between algorithmic decisions and the individuals they impact, which have shown to be prevalent in practical settings. Such obse
Despite prior research identifying several impactful scenarios where such dynamics can occur, including bureaucratic processes, social learning, recourse, and strategic behavior, extensive investigation of the long term effect of fairness methods remains limited. Initial studies have shown how enforcing static fairness constraints in dynamical systems can lead to unfair data distributions and may perpetuate or even amplify biases.
Additionally, the rise of powerful large generative models have brought at the forefront the need to understand fairness in evolving systems. The general capabilities and widespread use of these models raise the critical question of how to assess these models for fairness and mitigate observed biases within a long term perspective. Importantly, mainstream fairness frameworks have been developed around classification and prediction tasks. How can we reconcile these existing techniques (proprocessing, in-proc
Hima Lakkaraju