Assessing the Long-Term Impact of AI Bias Mitigation Techniques

This research works on assessing the long-term effects of AI bias mitigation techniques on different AI models and datasets like ResNet50 for image classification, BERT for sentiment analysis, and Logistic Regression for predictive policing, pre-, in- & post-processing techniques and fairness-aware deep learning models like FairGAN, Adversarial Debiasing & Fair Representation Learning. Results prove that initially there was a good reduction in bias but sustaining fairness over time is difficult because the bias re-appears with time, especially under dynamic environments. We underline that fairness is usually traded off for accuracy, and thus any technique should be evaluated according to multiple variables that consider various metrics and contextual factors. To tackle these challenges, we propose Dynamic Bias-Adaptive Learning (DBAL) a new framework which addresses these challenges by dynamically adjusting bias mitigation techniques based on observed fairness performance. This approach paves path for developing adaptive, scalable and responsible AI systems that are capable of sustaining fairness over time.