Imagine an AI that not only flags harmful content but also explains its reasoning, like a digital safety expert. That's the promise of SafetyAnalyst, a new framework designed to make AI content moderation more transparent, interpretable, and adaptable. Unlike current systems that often function as opaque black boxes, SafetyAnalyst builds a "harm-benefit tree." This tree maps out the potential consequences of an AI responding to a given prompt, considering who might be affected, the severity and likelihood of different outcomes, and even the potential benefits of providing a response. This detailed analysis then feeds into an algorithm that calculates a harmfulness score, which can be adjusted to align with specific safety preferences or community values. Researchers tested SafetyAnalyst by training a smaller, open-source model called SafetyReporter on data generated by larger language models. SafetyReporter learned to generate these harm-benefit trees and then classify prompts as harmful or benign. The results? SafetyReporter performed comparably to existing state-of-the-art models on a range of benchmarks, while offering significantly greater transparency. While SafetyAnalyst demonstrates an innovative approach to AI safety, challenges remain. Generating these detailed harm-benefit trees is computationally intensive. There are also open questions about how to best balance competing values and account for harms and benefits that are difficult to quantify. However, this research represents a critical step towards building more trustworthy and responsible AI systems.