Cohere, OpenAI, and AI21 Labs Develop Preliminary Best Practices for Large Language Models

Cohere, OpenAI, and AI21 Labs have developed a preliminary set of best practices applicable to any organization developing or deploying large language models. The available description presents the practices as broadly relevant to organizations that build or operate such models.

This framing suggests that the practices are intended as a general starting point rather than a product-specific policy or a closed technical specification. In that sense, the emphasis appears to be on shared expectations: how organizations can organize internal thinking about model development, deployment, and operational oversight when working with systems whose behavior can vary widely across tasks, users, and contexts. Because large language models are trained on broad corpora and can generate text or code in flexible ways, even preliminary guidance can serve as a common reference for teams that may otherwise use different terminology, workflows, or risk vocabularies.

The available materials also include titles referring to a hazard analysis framework for code synthesis large language models, evaluating large language models trained on code, and scaling laws for neural language models. These references point to recurring dimensions of model governance: anticipating failure modes, measuring behavior, and understanding how capabilities may change as model or data scale increases. Hazard analysis, in general, means identifying possible ways a system could produce harmful, unintended, or unreliable outcomes. Evaluation methods, in general, provide structured ways to test whether a model performs acceptably under defined conditions. Scaling laws, in general, describe the relationship between model size, compute, data, and observed performance, helping teams reason about how behavior may shift as systems grow.

Large language models are AI systems trained on large amounts of text or code, so hazard analysis, evaluation methods, and scaling laws are general areas used to examine how such systems behave as they are built and used. This background is relevant because best practices become more meaningful when they connect to these underlying questions. For example, a hazard analysis can help clarify where a model might create risk in code generation, such as producing misleading, insecure, or context-inappropriate output. Evaluation can help determine whether such outputs are frequent enough to require additional controls. Scaling insights can inform how much confidence an organization places in a model’s behavior before deployment, especially when the model is exposed to many users or tasks.

The materials do not provide further details about the content of the practices or the referenced frameworks. As a result, the immediate impact is likely to be conceptual: the listed organizations are associated with a preliminary reference point, and the accompanying titles signal that code synthesis, evaluation, and scaling are part of the broader discussion. For readers, the key takeaway is that responsible work on large language models is often organized around identifying hazards, testing behavior, and understanding capability trends, even when the available materials stop short of detailing concrete rules, procedures, or implementation steps.