Ethical Implications of Biases in AI and Machine Learning Algorithms

Due to data-driven decision-making, artificial intelligence (AI) systems are able to create "shadow profiles," which allow them to infer sensitive personally identifiable information, like gender, unnecessarily from indirect profiling sources without user consent. Even when those predictions can seem to be accurate, this experience raises serious ethical challenges and problems when an AI model’s predictions become biased against certain groups of individuals seeking to use a variety of platform purposes. In this work, we have investigated the shadow profile creation in ethical terms with the UCI Adult Dataset by building a model that accurately predicts gender. A Random Forest classifier was built to make gender predictions based on only the non-sensitive attributes in the dataset. With the use of SHAP (SHapley Additive exPlanations), we examined the features that influenced the model and potentially led to unequal performance across groups based on race. With Fairlearn implemented Exponentiated Gradient algorithm with demographic parity, racial bias deterioration was possible, but at the expense of decreased overall predictive accuracy of the model. This study adds to the argument for fairness-aware modeling and fairness-friendly explainability to help develop ethical AI systems.