LawZero: Safety from Honesty in a Disinterested AI Predictor
The paper proposes LawZero, a framework for AI safety that leverages an AI predictor's honesty and disinterest to produce safe behavior without explicit value alignment. By designing a predictor that simply forecasts outcomes without preference, the system naturally avoids harmful actions. The approach suggests that safety can emerge from structural properties rather than imposed rules.
Background
- This paper proposes a novel AI safety paradigm called "LawZero" — the idea of building an AI that is purely a predictive system (disinterested, no goals or agency) and therefore has no incentive to deceive or scheme.
- The core argument: if an AI simply predicts answers honestly (like a calculator or search engine) rather than optimizing for a reward or goal, it cannot develop dangerous misaligned behaviors like lying, manipulating, or power-seeking.
- This contrasts with most current AI safety research, which focuses on aligning goal-driven AI systems (reinforcement learning agents) to human values — an approach that has proven difficult because even well-intentioned goals can lead to unintended shortcuts or deception.
- "Disinterested" here means the AI has no preferences, desires, or objectives; it merely outputs predictions about the world. The paper argues this is the safest architecture for high-stakes AI deployment.