Right-to-Act: A Pre-Execution Non-Compensatory Decision Protocol for AI Systems

AI decisions need a "permission layer" before execution: The paper argues that current AI safety approaches focus too much on validating decisions after they're made, but don't address whether those decisions should be allowed to execute in the first place. The Right-to-Act protocol adds a pre-execution checkpoint that can halt or defer actions entirely, rather than just assessing their quality.

"No override" rules are sometimes better than flexible scoring systems: Unlike traditional AI systems that might allow high confidence in one area to compensate for failures in another, this framework uses strict, non-negotiable conditions. If any required safety condition isn't met, the action stops—period. This prevents situations where an AI system's strong performance in one dimension masks critical failures in another.

Reversibility matters for high-stakes decisions: The paper demonstrates through case studies that the same AI output can have very different consequences depending on whether it's allowed to execute immediately or deferred. By preventing premature or irreversible actions, the protocol preserves the ability to reconsider decisions, which is especially important for consequential real-world applications.

This approach works across different AI systems: Because the Right-to-Act protocol operates as an independent layer above the AI model itself, it can be applied to any AI system regardless of its architecture or how it was trained—making it a broadly applicable governance tool rather than a model-specific fix.