Towards AI-$45^{\circ}$ Law: A Roadmap to Trustworthy AGI

A structured framework is needed to balance AI safety with capability development. The paper proposes the "AI-45° Law" as a guiding principle and introduces the "Causal Ladder of Trustworthy AGI"—a three-layer framework (Approximate Alignment, Intervenable, and Reflectable layers) designed to systematically address safety challenges while allowing AI systems to advance in capability.

Trustworthy AGI requires five progressive levels of assurance. The framework defines distinct trustworthiness levels—perception, reasoning, decision-making, autonomy, and collaboration trustworthiness—that represent increasingly complex aspects of AI systems, suggesting that safety measures must evolve as AI capabilities grow.

Current safety proposals lack comprehensive guidance. Despite various safety assurance proposals and warnings about extreme risks, the paper identifies a critical gap: there are no comprehensive guidelines that adequately balance the dual goals of ensuring AI safety and enabling beneficial AI capabilities, particularly for high-autonomy or safety-critical applications.

Governance measures are essential alongside technical solutions. The paper emphasizes that trustworthy AGI development requires not only technical frameworks but also "a series of potential governance measures," indicating that responsible AI development cannot rely on technical approaches alone.

Causality-inspired design can improve AI trustworthiness. By adapting Judea Pearl's "Ladder of Causation" concept, the framework suggests that understanding causal relationships—rather than just correlations—can help create AI systems that are more interpretable, controllable, and aligned with human values.