PXGen: A Post-hoc Explainable Method for Generative Models

Explainability for generative AI is underdeveloped: The paper identifies that while explainable AI (XAI) has advanced significantly for other AI systems, "research in XAI for generative models remains relatively scarce," highlighting a gap in our ability to understand how popular generative AI tools actually work.

Good explanations need multiple qualities: The paper establishes that effective XAI methods should meet dual criteria—explanation quality (faithfulness, plausibility, completeness, and personalization) and system design principles (reliability, resilience, verifiability, and transparency)—suggesting that responsible AI requires more than just making outputs understandable.

Post-hoc explanation methods offer practical solutions: PXGen demonstrates a "post-hoc" approach that can be applied to existing generative models without rebuilding them, making explainability more feasible for AI systems already in use rather than requiring redesign from scratch.

Customizable explanations serve different stakeholder needs: The method allows users to customize "Anchor sets" and evaluation criteria according to their specific purposes and requirements, recognizing that different stakeholders (regulators, developers, end-users) may need different types of explanations.

Example-based explanations improve accessibility: By providing "example-based explanation methods," PXGen makes AI decision-making more interpretable to non-technical audiences, supporting broader accountability and trust in generative AI systems.