MoralMod: an Auditable & Steerable Moral Module for AI Systems

As AI systems transition from one-shot tools to persistent agents that interact over longer and longer time horizons, they will increasingly make decisions that could carry moral consequences for the humans they interact with. It is crucial that we develop interpretability methods to understand what went wrong when these systems make choices that conflict with their users' values, and steering methods to guide agents toward decisions that align with those values. We introduce a preliminary version of MoralMod, an AI system that can make interpretable and steerable moral choices in any context that can be expressed as natural language. The system takes a text-based vignette as input, builds an editable structured representation of that input, turns the representation into executable code, and finally outputs a value for each possible action that encodes how good or bad it would be to take that action. This system is both auditable and steerable, increasing safety guarantees and allowing users or organizations to iteratively improve the model's alignment with their values.