Best Practices for Deploying Language Models

LLM deployment principles were compiled based on experience providing API services, but they can also be applied to open-source releases or internal use. Since commercial use and safety issues are still new and constantly changing, the principles and practices will continue to be updated together with the industry.

• Blocking misuse: Usage guidelines and terms of service should prohibit harms such as spam, fraud, and manufacturing fake public opinion, and should also explicitly block high-risk use cases such as classifying people based on protected characteristics. This requires rate limit, content filtering, pre-approval for production access, and monitoring for signs of anomalies.
• Mitigating unintended harm: Comprehensive evaluations should be used to check model limitations, reduce bias in training data, and lower unsafe behavior through human feedback-based training.
• Documenting limitations: Weaknesses that are difficult to fully eliminate, such as bias or the possibility of generating security-vulnerable code, should be disclosed, along with safety guidelines specific to each model and use case.

Building teams with diverse backgrounds and broadly incorporating the opinions of external stakeholders is also key. Respecting supply-chain labor by allowing labelers to refuse certain tasks, and publicly sharing lessons learned from safety and anti-misuse efforts, helps drive iterative improvement across the industry.

Anthropic, Google, Microsoft, and Stanford CRFM, among others, expressed support for these principles, and emphasized the need for more detailed norms involving broader participation from academia, industry, and civil society.