Responsibility: An Example-based Explainable AI approach via Training Process Inspection

Training examples are more intuitive explanations than technical features: The paper's user study found that showing people the actual training examples an AI learned from ("this is what I learned") was more understandable than existing XAI methods, particularly for users without AI background—addressing a key gap in making AI explainability accessible to the general public.

Example-based explanations improve human decision-making: When users were shown which training example most influenced an AI's decision, they made more accurate judgments about whether to trust that decision, suggesting this approach could help people make better choices when relying on AI systems.

The method works across different types of problems: The researchers tested their "Responsibility" approach on multiple domains (not just one narrow application), indicating the technique has broad potential applicability beyond a single use case.

AI systems can become more transparent by examining their own training process: Rather than trying to reverse-engineer how a trained AI works, this method looks back at the training data to explain decisions—offering a practical way to make AI decision-making more auditable and understandable to stakeholders.

Better explanations benefit both humans and other AI systems: The approach improved accuracy not only for human users trying to understand AI decisions, but also when other machine learning models were trained to evaluate the explanations, suggesting clearer AI reasoning has cascading benefits.