Radical Interpretability Framework for AI Beliefs and Desires.

As AI systems become more complex and autonomous, understanding their internal workings, particularly their "beliefs" and "desires," is critical for safety and trust. This work proposes a novel framework for interpreting AI as agents, combining philosophical concepts of radical interpretation with the practical tools of mechanistic interpretability. The central challenge is to deduce an AI's internal states—its beliefs, desires, and the meanings it assigns—from its computational behavior. The framework establishes criteria for both representationalist and interpretationist approaches to AI interpretability, linking them to tests that current interpretability methods can perform. A key insight is the holistic nature of these attributions: beliefs, desires, and their underlying propositional structure are interdependent and cannot be analyzed in isolation. This holism is especially relevant for AI, which may operate with concepts fundamentally different from human ones. However, this interdependence also provides leverage, as a system's attitudes constrain its propositional structure, and mechanistic interpretability can help measure both, offering a path to reliably detect issues like deception in AI systems.