Some Critical and Ethical Perspectives on the Empirical Turn of AI Interpretability

Interpretability tools can paradoxically hinder ethical AI: The study found that when AI explanations are produced through empirical methods, they tend to have "low denunciatory power"—meaning they're less likely to reveal or highlight ethical problems. This creates a false sense of security, where companies may believe their AI systems are ethical simply because they can explain them, when explanations actually mask underlying issues.

One-size-fits-all explanations don't work: The research demonstrates that how well an AI explanation resonates depends heavily on who receives it, including factors like the recipient's gender and education level. This means a single explanation deemed "interpretable" by one group may fail to communicate ethical concerns to another, undermining the goal of transparent AI across diverse populations.

Self-regulation through explanations is insufficient: The paper argues that allowing companies to voluntarily produce their own AI explanations—the current "liberal" approach—is not adequate for addressing ethical concerns in AI systems. The flexibility of this approach enables companies to select explanations that downplay rather than highlight problems.

AI ethics requires stronger external oversight: The authors suggest that without more rigorous external regulation, the AI industry's current approach to interpretability will continue to fall short of genuine ethical accountability. The alternative of further liberalizing explanations would likely make this problem worse.