The ideal AI safety moderation system would be both structurally interpretable (so its decisions can be reliably explained) and steerable (to align to safety standards and reflect a community’s values), which current systems fall short on. To address this gap, we present SAFETYANALYST, a novel AI safety moderation framework. Given an AI behavior, SAFETYANALYST uses chainof-thought reasoning to analyze its potential consequences by creating a structured “harm-benefit tree,” which enumerates harmful and beneficial actions and effects the AI behavior may lead to, along with likelihood, severity, and immediacy labels that describe potential impact on any stakeholders. SAFETYANALYST then aggregates all harmful and beneficial effects into a harmfulness score using fully interpretable weight parameters, which can be aligned to particular safety preferences. We applied this conceptual framework to develop, test, and release an opensource LLM prompt safety classification system, distilled from 18.5 million harm-benefit features generated by frontier LLMs on 19k prompts. On a comprehensive set of prompt safety benchmarks, we show that SAFETYANALYST (average F1=0.81) outperforms existing LLM safety moderation systems (average F1<0.72) on prompt safety classification, while offering the additional advantages of interpretability, transparency, and steerability.

Learning structures that effectively abstract decision policies is key to the flexibility of human intelligence. Previous work has shown that humans use hierarchically structured policies to efficiently navigate complex and dynamic environments. However, the computational processes that support the learning and construction of such policies remain insufficiently understood. To address this question, we tested 1,026 human participants on a decision-making task where they could learn, transfer, and recompose multiple sets of hierarchical policies. We propose a novel algorithmic account for the learning processes underlying observed human behavior. We show that humans rely on compressed policies over states in early learning, which gradually unfold into hierarchical representations via meta-learning and Bayesian inference. Our modeling evidence suggests that these hierarchical policies are structured in a temporally backward, rather than forward, fashion. Taken together, these algorithmic architectures characterize how the interplay between reinforcement learning, policy compression, meta-learning, and working memory supports structured decision-making and compositionality in a resource-rational way.

Computational cognitive modeling is an important tool for understanding the processes supporting human and animal decision-making. Choice data in decision-making tasks are inherently noisy, and separating noise from signal can improve the quality of computational modeling. Common approaches to model decision noise often assume constant levels of noise or exploration throughout learning (e.g., the softmax policy). However, this assumption is not guaranteed to hold – for example, a subject might disengage and lapse into an inattentive phase for a series of trials in the middle of otherwise low-noise performance. Here, we introduce a new, computationally inexpensive method to dynamically estimate the levels of noise fluctuations in choice behavior, under a model assumption that the agent can transition between two discrete latent states (e.g., fully engaged and random). Using simulations, we show that modeling noise levels dynamically instead of statically can substantially improve model fit and parameter estimation, especially in the presence of long periods of noisy behavior, such as prolonged lapses of attention. We further demonstrate the empirical benefits of dynamic noise estimation at the individual and group levels by validating it on four published datasets featuring diverse populations, tasks, and models. Based on the theoretical and empirical evaluation of the method reported in the current work, we expect that dynamic noise estimation will improve modeling in many decision-making paradigms over the static noise estimation method currently used in the modeling literature, while keeping additional model complexity and assumptions minimal.