Bias & Fairness in AI Models: Developing Techniques to Reduce Bias in AI-Generated Decisions

With the proliferation of artificial intelligence (AI) in decision-making, the potential biases inherent in algorithms come into sharper relief - especially in high-stakes areas like criminal justice. AI models trained on real world data often inherit and magnify biases in that data, resulting in differences in predictions among demographic groups. In this work, we explore disparities in AI-generated decisions using the COMPAS recidivism dataset with respect to false positive and false negative rates among racial cohorts. To improve fairness, we leverage reweighing - a pre-processing bias mitigation technique - to obtain adjusted sample weights that enhance fairness at no cost of accuracy to the model. Findings reveal that without bias mitigation, false positive rates were significantly higher for both Caucasian and Asian individuals, while individuals grouped in the “Other” category showed a tendency for the highest false negatives, which results in an unfair treatment. Reweighing ameliorated the false negative rates in non-Caucasian groups significantly, which was an unjust tendency to classify patients as low-risk falsely. These fairness gains were made without sacrificing overall model accuracy or predictive performance, an important feature of Matthias et al.'s method. Overall, the study demonstrates the impact of bias mitigation techniques on biases in decisions made by the AI. But there are still some disparities - indicating room for improvements. Our method does not address all the existing problems in fairness in AI, future work will focus on investigating post-processing adjustments, alternative classifiers, explainability and other potential techniques to promote fairness in AI decision making.