Attributing Accountability in Multi-Stage AI Systems

A new research paper from arXiv addresses the challenge of attributing accountability in modern AI systems, which are developed through multiple stages including pretraining, fine-tuning, and alignment. The authors propose a framework to answer counterfactual questions about how model behavior would differ if a particular stage's updates were omitted. They introduce estimators that quantify stage effects without retraining, accounting for data and optimization dynamics. The work aims to trace model behavior back to specific development stages, raising critical questions about responsibility for success or failure.