OpenAI, Cohere, and AI21 Labs Best Practices for Deploying Language Models

OpenAI, Cohere, and AI21 Labs have established a preliminary set of best practices for organizations developing or deploying large language models (LLMs). These guidelines aim to mitigate the risks associated with powerful AI technology to ensure it augments human capabilities while minimizing harm.

Prohibiting Misuse of Language Models

LLM providers should implement strict usage guidelines and technical infrastructure to prevent the technology from being used for malicious purposes.

  • Usage Guidelines and Terms of Use: Providers must publish clear terms that prohibit material harm to society, individuals, and communities. This includes banning the use of LLMs for spam, fraud, or astroturfing. Additionally, guidelines should identify high-risk use-cases that require extra scrutiny or are entirely prohibited, such as the classification of people based on protected characteristics.
  • Enforcement Infrastructure: To ensure compliance with usage guidelines, providers should build systems such as rate limits, content filtering, application approval processes for production access, and monitoring for anomalous activity.

Mitigating Unintentional Harm

Because no preventative action can completely eliminate the potential for unintended harm, providers must proactively address model behavior and be transparent about remaining vulnerabilities.

  • Proactive Mitigation: Providers should employ comprehensive model evaluation to assess limitations and use techniques like learning from human feedback to minimize unsafe behavior. Efforts should also be made to minimize potential sources of bias within training corpora.
  • Documentation of Weaknesses: Providers must document known vulnerabilities, including the model's ability to produce insecure code or inherent biases. This documentation should include safety best practices specific to the model and its intended use-cases.

Stakeholder Collaboration and Ethical Labor

Responsible deployment requires diverse perspectives and industry-wide transparency to address how models operate in the real world.

  • Diverse Teams: Organizations should build teams with diverse backgrounds to ensure that the models do not reinforce biases or fail for specific groups of people.
  • Public Disclosure: Lessons learned regarding LLM safety and misuse should be shared publicly to enable cross-industry iteration and widespread adoption of safety standards.
  • Labor Standards: Providers must treat all labor in the LLM supply chain with respect. This includes maintaining high standards for the working conditions of those reviewing model outputs in-house and ensuring third-party vendors adhere to well-specified standards, such as allowing labelers to opt out of specific tasks.

Industry Support and Perspectives

The joint recommendation has received support from various AI organizations and research centers, emphasizing the importance of industry collaboration.

"While LLMs hold a lot of promise, they have significant inherent safety issues which need to be worked on. These best practices serve as an important step in minimizing the harms of these models and maximizing their potential benefits."

—Anthropic

"As large language models (LLMs) have become increasingly powerful and expressive, risk mitigation becomes increasingly important. We welcome these and other efforts to proactively seek to mitigate harms and highlight to users areas requiring extra diligence."

—John Bansemer, Director of the CyberAI Project and Senior Fellow, Center for Security and Emerging Technology(CSET)

"Google affirms the importance of comprehensive strategies in analyzing model and training data to mitigate the risks of harm, bias, and misrepresentation."

—Google

"To realize the promise of large language models, we must continue to collaborate as an industry and share best practices for how to responsibly develop and deploy them while mitigating potential risks."

— Microsoft

"The safety of foundation models, such as large language models, is a growing social concern. We commend Cohere, OpenAI, and AI21 Labs for taking a first step to outline high-level principles for responsible development and deployment..."

—Percy Liang, Director of the Stanford Center for Research on Foundation Models(CRFM)

Sources