A Framework for Inherently Safer AGI through Language-Mediated Active Inference

Safety built in from the start, not added later: The framework argues that current AI safety approaches—like trying to interpret what an AI is thinking after the fact or adjusting its reward system—have fundamental flaws. Instead, this paper proposes designing safety into AGI's core architecture from the beginning, similar to how safety features are engineered into bridges rather than patched afterward.

Natural language as a transparency tool: By using natural language to represent what the AI believes and what it values, humans can directly oversee and understand the system's reasoning. This approach aims to make AI decision-making more readable to people, rather than relying on complex technical explanations that only specialists can verify.

Separating "what the AI thinks" from "what the AI wants": The framework explicitly distinguishes between an AI's beliefs about the world and its preferences or goals. This separation in natural language could help prevent scenarios where an AI pursues goals in unexpected ways, since its objectives would be clearly stated and separately auditable from its factual understanding.

Modular design for safer scaling: The proposal uses a multi-agent system with modular structures, where safety constraints flow through the system's hierarchy. This compositional approach suggests that safety properties could be maintained as AI systems become more complex, rather than safety guarantees breaking down at larger scales.

Testable framework with concrete benchmarks: Rather than remaining theoretical, the authors propose validating their approach using the Abstraction and Reasoning Corpus (ARC) benchmark, providing a concrete way to test whether their safety mechanisms actually work in practice.