Position: Current Model Cards Are Insufficient for Downstream Governance of Open-Weight Foundation Models

What they did

1. Surveyed model cards, AUPs, and licenses of the top 500 most-downloaded open-weight foundation models on Hugging Face, using the NIST GenAI Risk Management Framework's safety keyword taxonomy
2. 99.6% of models had a model card, but only 75.2% included safety-specific fields, and only 21.2% had an explicit AUP referenced in the card
3. Documentation depth showed no meaningful link to download popularity (regression R2<0.01), and instead tracked which vendor made the model — Meta and Google scored higher, Qwen lower
4. Argued that common open-source licenses like Apache 2.0 and MIT are structurally unable to carry use-based restrictions, which weakens or overrides any accompanying AUP legally
5. Proposed a standardized 'safety card' template emphasizing model heritage (what upstream models/data shaped it) and operational safety testing, plus a redesigned OWFM-specific license that legally binds an AUP as an exhibit

Why it matters

This gives downstream developers and regulators concrete evidence that today's model cards are not sufficient for assessing real safety risk of open-weight models. It offers a practical blueprint — informational, normative, and legal layers — for anyone building or auditing open-weight model governance going forward.

Terms in this paper

• OWFM (open-weight foundation model) · an AI model released with public weights but without public training code or data
• model card · a standard document describing a model's capabilities, limitations, and risks
• AUP (acceptable use policy) · a normative document listing prohibited or restricted uses of a model
• OSL (open-source license) · licenses like Apache 2.0 or MIT originally designed for software code
• NIST GenAI Risk Management Framework · a U.S. government taxonomy for categorizing generative AI risks

Figures we cannot republish

• Figure 1: A Three-Layered Approach to Downstream Governance of Open-Weight Foundation Models. (Left) AS-IS: The current fragmented system, where model cards focus on performance rather than safety, AUPs are often missing or non-binding, and model licenses frequently conflict with stated usage restrictions. (Right) TO-BE: The proposed unified framework integrating three complementary layers for more resilient downstream safety control: informational (safety cards), normative (standardized AUPs), and legal (OWFM-tailored licenses).
• Figure 2: Prevalence of Safety-Related Keywords Across the Top 500 Downloaded Models. Each bar represents the number of models whose documentation contains at least one keyword from the corresponding safety category, based on the NIST GenAI Risk Management Framework taxonomy.
• Figure 3: Safety-Related Keyword Coverage by Vendor (Top 12 by Model Count). Stacked bars show the number of models in each keyword-count range (0, 1–5, 6–10, 11–20, 21+). Vendors such as meta (including Llama and facebook OPT models) and google exhibit higher keyword density, while Qwen models are concentrated in lower-frequency ranges, highlighting systematic differences in disclosure practices.
• Figure 4: Vendor-Level Clustering of Model Card Embeddings. Each point represents a model, positioned using PCA on sentence embeddings of model card text. Colors indicate developer affiliation, and point size reflects total governance-related keyword counts. k-means clustering yields several distinct developer-level documentation patterns, while a heterogeneous set of other developers remains broadly dispersed across the embedding space.
• Figure 5: Relationship Between Model Popularity and Governance Documentation Depth. Each point represents a model, with total downloads (millions) on the x-axis and governance-related keyword counts on the y-axis. Regression lines across multiple functional forms are not explanatory (R2<0.01), hinting that ecosystem utility does not predict documentation quality.
• Figure 6: Governance Artifact Flows Across the Top 500 OWFMs. The Sankey diagram traces each model across download rank, safety-keyword coverage, developer, license, and AUP status. Flow widths indicate model counts, showing that permissive OSLs, notably Apache 2.0 and MIT, lead to "no AUP", while custom licenses such as Llama and Gemma are associated with AUP presence.

Latest papers

• SWE-bench Science: Can Coding Agents Resolve Engineering Tasks in Science?AI coding agents were tested on fixing real scientific software, and even the best one failed more than half the timearXiv:2608.197992026-08-21
• FlashPrefill V2: Block-Sparse Prefill Attention for Long-Context LLM ServingMaking sparse attention fast enough and accurate enough for real LLM serving, not just papersarXiv:2608.197582026-08-21
• PolicyGuide: From Guarding One Action to Guiding the Whole Workflow for Policy-Compliant LLM AgentsMaking customer-service AI agents follow the whole procedure, not just avoid one bad actionarXiv:2608.198612026-08-21
• EXIMO: VLM Guided Exploration of VLA PoliciesTeaching a robot new chores without human teleoperation, by letting a chatty AI supervise itarXiv:2608.198912026-08-21
• EnvHarness: Awakening Static Worlds for Agent LearningInstead of building new training worlds from scratch, this work adds a plug-in layer that reshapes existing ones around each agent's actual weaknessesarXiv:2608.198802026-08-21
• Bounded Sovereignty and the Control Tax: Pricing AI Oversight When the Deployer Does Not Own the ModelCompanies that rent AI instead of owning it can only do half of AI safety oversightarXiv:2608.192162026-08-21
• Beyond Imitation: Filtering On-Policy Distillation by Reasoning ProgressA fix for AI models that get penalized by their teacher even when they're reasoning correctlyarXiv:2608.194082026-08-21
• PersonalBench: Measuring the Authorship Gap in LLM PersonalizationAI can be prompted to write 'like someone,' but its own voice never fully disappearsarXiv:2608.197462026-08-21

Latest from METAL MEDIA

• TUIDE debut with 'SUN KISS,' an Iceland-shot K-pop road movie
• NCT 127 drop BLINGY, their first full album in two years