Causal Responsibility Attribution for Human-AI Collaboration

  • 2024-11-05 17:17:45
  • Yahang Qi, Bernhard Schölkopf, Zhijing Jin
  • 0

Abstract

As Artificial Intelligence (AI) systems increasingly influencedecision-making across various fields, the need to attribute responsibility forundesirable outcomes has become essential, though complicated by the complexinterplay between humans and AI. Existing attribution methods based on actualcausality and Shapley values tend to disproportionately blame agents whocontribute more to an outcome and rely on real-world measures ofblameworthiness that may misalign with responsible AI standards. This paperpresents a causal framework using Structural Causal Models (SCMs) tosystematically attribute responsibility in human-AI systems, measuring overallblameworthiness while employing counterfactual reasoning to account for agents'expected epistemic levels. Two case studies illustrate the framework'sadaptability in diverse human-AI collaboration scenarios.

 

Quick Read (beta)

loading the full paper ...