Towards Fair and Explainable AI using a Human-Centered AI Approach

Human judgment matters more than performance metrics alone: The paper demonstrates that fairness, trust, and explainability are subjective and context-dependent qualities that don't automatically correlate with conventional performance metrics, suggesting AI systems need to be evaluated on human-centered criteria beyond accuracy.

Explanations build appropriate trust in AI systems: The first research project found that providing local model explanations to human trainers (crowd workers) supports "trust calibration," meaning people develop more accurate and appropriate levels of confidence in the AI system rather than either over-trusting or under-trusting it.

Interactive tools can help identify hidden biases: The D-BIAS tool demonstrates that visual, human-in-the-loop approaches can help identify and mitigate social biases in datasets, while simultaneously enhancing trust and accountability—showing that bias detection benefits from human involvement rather than being purely automated.

Bias exists in foundational AI components: WordBias reveals that biases against specific groups (such as women or intersectional groups like Black Muslim females) are embedded in pre-trained word embeddings, which are widely used building blocks in AI systems, indicating bias problems start early in the AI development pipeline.

Human-centered AI requires transparency and control: The paper's core argument is that empowering people through greater transparency and human control over AI systems is essential for addressing fairness and accountability concerns that fully automated approaches cannot solve alone.