A New Framework for Open Source AI

Defining Openness in the Age of AI

The debate over open versus closed AI has moved from academic circles into the realm of global policy, with significant implications for innovation, safety, and competition. Mozilla and the Columbia Institute of Global Politics have taken a step toward resolving this debate by introducing a framework that redefines what openness means in the context of AI systems. This framework is not a simple binary choice between open and closed, but rather a nuanced approach that considers the various components of AI development.

Breaking Down the AI Stack

The framework breaks down the AI stack into its core components—data, code, model weights, documentation, and more—each of which can be open or closed independently. This granular approach allows for a more precise evaluation of how openness at each layer affects the ability of different stakeholders to inspect, use, and build upon a system. This specificity is crucial, as many current policy discussions treat open source AI as a monolithic concept, ignoring the nuanced trade-offs that arise from varying levels of openness.

Openness as a Gradient

The framework treats openness as a gradient rather than a binary. For example, a system may be open on model weights but closed on training data, or open on code but lack detailed documentation on evaluation processes. This approach allows for a more accurate description of the actual state of openness in AI systems, rather than relying on vague labels.

Safety Beyond the Model

The framework also emphasizes that safety cannot be assessed solely at the model level. It highlights the importance of the surrounding system, including deployment environments, safeguards, moderation layers, and governance structures. These elements collectively shape the real-world risk profile of an AI system, making it essential to consider them alongside the model itself.

A Shared Vocabulary for Policymakers and Developers

One of the key contributions of the framework is that it provides a common vocabulary for developers, researchers, and regulators to evaluate trade-offs on a case-by-case basis. Rather than prescribing a single “correct” level of openness for all systems, the framework offers a flexible approach that allows for tailored decisions based on specific contexts and needs.

Collaborative Effort

The paper was co-authored by a diverse group of individuals and organizations from academia, industry, and civil society. This cross-section of perspectives reflects a genuine effort to move the development of open systems forward, grounded in years of discussion and collaboration. The framework represents a milestone in the ongoing mission to make AI systems accessible to a broader range of stakeholders, rather than being controlled by a few.

Source: Mozilla. Analysis based on publicly available information; no hands-on testing is claimed.