Towards AI-45° Law: A Roadmap to Trustworthy AGI
Introduction: The Era of "Crippled AI"
As artificial intelligence continues its rapid ascent, driven by scaling laws and unprecedented compute, a critical imbalance has emerged. We are currently witnessing the rise of "Crippled AI"—systems possessing immense capability but lacking proportional safety and trustworthiness. Current safety measures are often reactive, fragmented, and easily bypassed, leading to powerful yet fragile systems that risk undermining public trust and posing existential threats. To address this, a new theoretical framework is required to ensure safety is an architectural pillar of Artificial General Intelligence (AGI) rather than a post-hoc addition.
The AI-45° Law: A Guiding Principle
The authors propose the AI-45° Law as a fundamental coordinate system for AGI development. In this model, the x-axis represents "AI Capability" while the y-axis represents "AI Safety." The ideal trajectory for development follows a 45-degree diagonal, where safety advancements keep pace with capability gains.
To operationalize this law, the framework defines two critical boundaries:
- The Yellow Line: Represents early warning indicators where capability begins to outpace safety, signaling a need for immediate intervention.
- The Red Line: Marks the threshold of catastrophic risk, including autonomous replication, deception, or irreversible loss of human control.
The Causal Ladder of Trustworthy AGI
Inspired by Judea Pearl’s research into causality, the authors introduce the Causal Ladder of Trustworthy AGI. This three-tiered technical framework moves beyond simple data-driven alignment toward sophisticated, counterfactual self-reflection.
The ladder consists of three progressive layers:
- Approximate Alignment Layer (Association): This foundational layer focuses on "What is it?" It utilizes data-driven methods like Supervised Fine-Tuning (SFT) and machine unlearning to fit models to human values.
- Intervenable Layer (Intervention): This layer addresses "What if I do X?" It involves mechanistic interpretability and human-in-the-loop feedback (such as RLHF and RLAIF) to monitor and modify a model's reasoning processes in real-time.
- Reflectable Layer (Counterfactuals): The highest tier focuses on "What would have happened if I chose differently?" By incorporating world models and self-reflection, the AI can deliberate on its own values and simulate the ethical consequences of its actions before executing them.
The framework further distinguishes between Endogenous Trustworthiness (safety mechanisms inherent to the system's design) and Exogenous Trustworthiness (external assurance, governance, and moderation).
Assessing the Current Landscape
Currently, state-of-the-art models like GPT-4 and Llama-3 primarily occupy the Approximate Alignment and Intervenable layers. While significant progress has been made in "Perception Trustworthiness," the field remains underdeveloped in the higher-order domains of Autonomy and Collaboration Trustworthiness required for true AGI.
A notable milestone is the recent OpenAI-o1 (preview), which the authors classify as a basic form of the Reflectable Layer. By integrating policy review and safety verification directly into its reasoning chain, it represents a shift toward the counterfactual reasoning necessary for Level 3 trustworthiness.
Conclusion and Future Directions
The AI-45° Law provides a proactive roadmap for the safe development of AGI. However, transitioning from theory to practice involves significant technical hurdles, particularly in implementing the Reflectable Layer and ensuring safety in multi-agent environments. The authors advocate for treating AI safety as a Global Public Good, necessitating multi-stakeholder lifecycle management to ensure that the pursuit of capability never crosses the "Red Line" of catastrophic risk.
Summary updated