Model cards may fall short for governing open-weight AI

A new arXiv position paper argues that current model cards are insufficient for governing open-weight foundation models after release. Model cards are documents that describe a model’s training, evaluation, intended uses, and limitations.

The paper’s concern is that open-weight models can be adapted, merged, or deployed in settings far removed from their original release context. A static document may not give downstream developers, auditors, or policymakers enough information to assess changed risks.

The argument does not reject transparency documentation. Instead, it suggests that governance needs richer, more durable mechanisms when model weights are widely available. That distinction matters as open models become a larger part of commercial and public-sector AI systems.