Meaningful human control: actionable properties for AI system development

Define clear boundaries for AI decision-making: AI systems should have explicitly defined domains specifying which morally significant situations they can handle autonomously versus which require human intervention. The paper illustrates this through automated vehicles and hiring systems, where the scope of AI authority must be clearly established upfront.

Ensure humans and AI "speak the same language": For meaningful control to work, humans and AI systems need compatible ways of representing information and decisions. Without this mutual understanding, humans cannot effectively oversee or intervene in AI actions, creating a control gap despite good intentions.

Match responsibility to actual control: Humans should only be held responsible for AI system outcomes to the extent they actually have the ability and authority to control those outcomes. The paper emphasizes that assigning blame to someone without real power over a system's decisions is both unfair and ineffective for accountability.

Create explicit accountability mechanisms: Rather than relying on general principles, the paper calls for concrete, engineered links between human decision-makers and AI system actions. This bridges the gap between philosophical concepts of responsibility and practical engineering requirements that developers can actually implement.

Address the responsibility gap in autonomous systems: The paper identifies a critical problem where autonomous AI systems can create situations where no one can be properly held accountable for harmful outcomes—a gap that meaningful human control is designed to prevent through these four actionable properties.