From Protocols to Evidence: Bounded Claims for AI in Service of the Common Good

AI claims often outpace evidence: The paper identifies a significant gap between what organizations claim AI systems accomplish (improving decisions, broadening access, reducing harm) and what their evaluations actually support, highlighting the need for more rigorous, bounded claims about AI capabilities.

Deployment can mask or worsen existing problems: The authors propose a "rupture test" to evaluate whether AI deployment actually fixes institutional failures or simply conceals them behind improved performance metrics—comparing AI solutions against existing human practices and non-AI alternatives to ensure genuine improvement rather than superficial fixes.

Ownership and institutional control matter as much as performance: Responsible AI requires examining who owns the infrastructure, controls how it's used, and bears responsibility for consequences. Favorable evaluation results alone cannot establish that a system respects human dignity or has moral legitimacy without addressing these power structures.

Evidence has limits in establishing legitimacy: The paper distinguishes between "evidence-bounded deployment" (limiting claims to what data actually supports under specific conditions) and "measurement-bounded governance" (identifying constraints that good results cannot override), recognizing that strong performance metrics don't automatically justify deployment on ethical grounds.

A framework for responsible AI deployment: RISE AI proposes examining systems across four dimensions—Responsibility, Inclusivity, Safety, and Empowerment—to move beyond technical evaluation toward assessing whether AI serves the common good while respecting human dignity.