The Governance Gap in Open-Weight Models
Current model cards—the standard documentation format for AI models—are designed for static, closed-source systems where the provider maintains control over the model's environment and usage. The authors argue that this framework is fundamentally broken for open-weight foundation models. Because these models are distributed, modified, and fine-tuned by downstream users, the original model card becomes obsolete almost immediately upon release. The current documentation fails to account for the 'governance drift' that occurs when a base model is transformed into a specialized application.
Moving Toward Dynamic Accountability
To address this, the paper proposes a shift from static documentation to a more robust, multi-layered governance framework. The authors highlight that downstream users often lack the resources or expertise to perform the rigorous safety evaluations that original developers conduct. Therefore, governance must evolve to include:
- Provenance Tracking: Establishing a clear lineage of modifications, ensuring that fine-tuned versions can be traced back to their base models and original safety constraints.
- Modular Documentation: Moving away from a single 'card' toward a living document that updates as the model is adapted, allowing for transparency in how specific fine-tuning processes might have introduced new risks or biases.
- Standardized Evaluation Protocols: Creating shared, machine-readable benchmarks that can be applied consistently across different versions of a model, rather than relying on subjective or non-comparable reporting from various downstream developers.
By treating model governance as a continuous process rather than a one-time disclosure, the authors suggest that the industry can better manage the risks associated with the proliferation of open-weight systems while maintaining the benefits of open innovation.