Towards Unified Attribution in Explainable AI, Data-Centric AI, and Mechanistic Interpretability

Attribution methods across AI interpretability share common underlying techniques: Feature attribution (explaining which inputs matter), data attribution (identifying influential training examples), and component attribution (understanding internal model mechanisms) all rely on similar technical approaches like perturbations, gradients, and linear approximations—they just apply these techniques to different aspects of AI systems.

A unified framework can improve AI transparency and trustworthiness: By recognizing these three attribution approaches as part of one coherent system rather than separate fields, researchers can develop better evaluation standards and shared concepts that make it easier to understand why AI systems make specific decisions, which is essential for responsible deployment.

This unified view enables practical applications for AI governance: The paper identifies concrete applications including model editing (fixing problematic behaviors), steering (guiding model outputs), and regulation (ensuring compliance)—suggesting that better interpretability frameworks could help organizations actually control and govern AI systems in practice.

The fragmented current landscape creates inefficiencies in AI safety research: Different communities studying feature, data, and component attribution work largely independently with inconsistent terminology and methods, meaning insights from one area aren't easily transferred to others, slowing progress on understanding AI behavior comprehensively.

Cross-attribution innovation could address shared interpretability challenges: By treating these three attribution types as interconnected rather than separate, researchers can tackle common problems (like evaluation criteria and validation methods) more effectively and develop new interpretability techniques that leverage insights across all three domains.