Mitigating Legibility Tax with Decoupled Prover-Verifier Games

Checkability matters for AI safety: As AI systems become more powerful, we need ways for humans and simpler AI systems to verify their outputs are correct—a challenge the paper calls the "legibility tax" problem.

There's a trade-off between accuracy and explainability: Current methods that make AI outputs easier to check suffer from reduced accuracy compared to models optimized purely for correctness, suggesting a fundamental tension between these two goals.

Separating tasks can improve both accuracy and checkability: The proposed solution trains one model to find correct answers first, then trains a separate "translator" model to convert those answers into a form that's easier to verify—achieving better results on both dimensions than trying to do both simultaneously.

Game theory provides a framework for trustworthy AI: The decoupled prover-verifier game approach uses mathematical equilibria to ensure the translator model faithfully represents the original answer while making it checkable, rather than relying on the model to balance competing objectives.

This addresses a real bottleneck in AI oversight: As AI systems outpace human ability to understand their reasoning, having reliable methods to verify outputs becomes essential for maintaining human control and trust in AI systems.