Researchers argue model cards alone can't govern open AI models
Landmark paper finds model cards fail to capture safety risks of OWFMs.
A team of researchers led by Sungwon Chae from Seoul-based institutions has published a position paper arguing that existing model cards are insufficient for the downstream governance of open-weight foundation models (OWFMs). The paper, accepted at ICML 2026, analyzed 500 model cards on Hugging Face and found critical gaps in how safety risks, alignment provenance, and model heritage are communicated. The authors contend that current open-source licenses (OSLs) are ill-suited for OWFMs and may undermine enforceable acceptable use policies (AUPs).
The researchers propose evolving model cards, AUPs, and licenses into integrated safety artifacts that address informational, normative, and legal dimensions. They highlight that standard model cards lack the granularity to inform downstream developers and users about unique safety challenges posed by OWFMs. The paper calls for a multi-layered governance framework, positioning this shift as essential for responsible AI deployment.
- Analyzed 500 Hugging Face model cards and found critical gaps in safety transparency for OWFMs.
- Proposes integrating model cards with AUPs and licenses into a unified governance framework.
- Argues standard open-source licenses are inadequate for governing open-weight foundation models.
Why It Matters
This research could redefine transparency standards for open AI models and reduce safety risks in downstream deployments.