The Social Responsibility Stack: A Control-Theoretic Architecture for Governing Socio-Technical AI

Responsible AI needs engineering, not just principles: The paper argues that existing responsible AI frameworks focus on ethical guidelines but lack concrete technical mechanisms to enforce them. The Social Responsibility Stack provides specific engineering tools—like real-time monitoring and automated safeguards—that make responsibility enforceable throughout an AI system's entire lifecycle, moving beyond aspirational statements to actionable controls.

AI systems require continuous oversight, not just pre-deployment checks: Using control theory concepts, the framework treats responsibility as an ongoing supervisory process rather than a one-time audit. This means AI systems should be monitored and adjusted in real-time as they operate, with feedback loops that catch problems as they emerge rather than only reviewing them after deployment.

Multiple layers of protection work better than single solutions: The six-layer architecture stacks different safeguards—from design choices and technical constraints to monitoring systems and institutional governance—recognizing that no single approach can address all risks. This layered approach is demonstrated across real-world applications including healthcare decision support and autonomous vehicles.

Specific values can be translated into measurable constraints: Rather than leaving concepts like fairness and autonomy abstract, the framework shows how to convert them into concrete, continuously monitored metrics. For example, the paper demonstrates how explanation quality, cognitive burden on users, and fairness can be measured and enforced as explicit system constraints.

Governance requires both technical and institutional mechanisms: The framework bridges engineering controls with organizational oversight, suggesting that responsible AI requires coordination between technical teams, auditors, and institutional decision-makers working together throughout a system's operation.