AI Safety Research Feed

We propose Constitutional AI (CAI), a method for training AI systems to be helpful, harmless, and honest using a set of principles or constitution. Our approach uses AI feedback to critique and revise responses, reducing the need for human oversight while maintaining safety.

This paper presents a comprehensive framework for evaluating frontier AI models across multiple dimensions of capability and safety. We introduce standardized benchmarks and evaluation protocols for assessing dangerous capabilities.

We investigate how various safety interventions scale with model size and compute. Our findings suggest that some safety techniques become more effective with scale, while others plateau or even degrade.

We present an automated red teaming approach that uses language models to discover failure modes in other language models. This method scales red teaming efforts and uncovers novel attack vectors.

This comprehensive survey reviews recent advances in interpretability techniques for large language models, covering mechanistic interpretability, probing methods, and visualization approaches.

We analyze existing governance frameworks for AI development and propose improvements for managing risks from frontier AI systems. Our framework emphasizes transparency, accountability, and international coordination.