Quantifying Theoretical AI Alignment Guarantees: Receiver-Utility Bounds in Bayesian Persuasion

The paper presents a theoretical analysis of AI alignment through the lens of Bayesian persuasion, focusing on the information flow from an AI agent to a human receiver. It establishes a utility bound, showing that the maximum utility attainable by the receiver when the AI optimizes a misaligned objective is at most 1.5 times the utility derived from the prior alone, with specific improvements noted for certain priors. This work is significant for practitioners as it quantitatively characterizes the limitations of information transfer in misaligned AI systems, informing strategies for better aligning AI objectives with human decision-making.