Accounts, Accountability and Agency for Safe and Ethical AI

Explainability alone isn't enough: The paper demonstrates that simply making AI systems explainable doesn't automatically make them safe or ethical—what matters is whether explanations are actually usable and understandable to the people who need to work with them in real-world settings.

Context matters for accountability: Using a mammography screening example, the authors show that accountability in AI-assisted decision-making is built through ongoing interaction between humans and systems. Generic explanations fail when they don't account for how teams actually collaborate and make decisions together.

Current XAI tools have practical limitations: The paper reveals that users struggled to interpret AI tool behavior even when explanations were provided, suggesting that current explainable AI approaches may not adequately address how people actually need to understand and trust these systems in high-stakes domains like healthcare.

Agency and accountability are intertwined: The research emphasizes that for AI to be truly ethical, we need to understand not just what decisions are made, but who is responsible for them—and this requires examining how humans and AI systems share decision-making authority in practice.

We need better research frameworks: The authors call for a new research agenda focused on understanding XAI requirements in real collaborative contexts, rather than treating explainability as a purely technical problem separate from how people actually work.