A collection of principles for guiding and evaluating large language models

37 core principles were identified across seven categories to guide and evaluate how large language models reason and behave, synthesized from research on reasoning, explainability, AI safety, ethics, and regulatory guidelines—providing a structured framework for responsible AI development beyond single-issue approaches.

Transparency and truthfulness remain critical gaps: The paper identifies challenges with LLMs' ability to explain their reasoning, verify information accuracy, and operate transparently, suggesting these capabilities need substantial improvement before deployment in high-stakes applications.

Principles can serve multiple practical purposes including real-time monitoring of models during use, improving model training, and supporting human evaluation—meaning a shared set of principles could help organizations implement responsible AI practices across different stages of model development and deployment.

Expert consensus on principle importance varies, as indicated by the expert survey component, highlighting that different stakeholders may prioritize different aspects of responsible AI (such as ethics vs. utility), which has implications for how organizations should tailor AI governance to their specific contexts and values.

Ethical alignment and robustness require ongoing attention: The paper emphasizes that LLMs face persistent challenges in ethical alignment and security/robustness, indicating that responsible AI requires continuous monitoring and refinement rather than one-time fixes.