As artificial intelligence becomes increasingly woven into the fabric of public health, the need for transparency and accountability in AI systems is more urgent than ever. We are already witnessing the impact of AI in public health surveillance, predictive analytics for outbreaks, resource allocation modeling, and even behavioral health interventions. But behind each algorithm lies a question we must ask: Do we truly understand how this model works, what its limits are, and where it might fail? This is where AI model cards come in.
Think of a model card as the “nutrition label” for an AI system. Just like food labels help consumers understand what they’re putting into their bodies, model cards offer users and stakeholders a way to understand what they’re relying on when they use an AI model. These documents explain how a model was trained, what data it learned from, its strengths and weaknesses, and where caution is advised. In essence, a model card is a standardized way to communicate the inner workings and outer limitations of machine learning models — especially those used in sensitive contexts like public health (Hugging Face, 2024).
Imagine a public health department using an AI tool to predict which neighborhoods are at highest risk for flu outbreaks. On the surface, the tool might look robust — the interface is sleek, predictions seem timely, and early results are promising. But without documentation, the health team might not know that the model underperforms in rural areas because the training data was drawn largely from urban health systems. Or that its predictions weaken in populations with limited access to health care. A model card would make this visible, helping teams adjust their expectations, apply appropriate safeguards, or consider retraining the model for better equity.
OpenAI’s GPT-4, for instance, comes with a comprehensive “system card” that explains not just what the model can do, but also how it might be misused — for misinformation, bias amplification, or unintended decision-making harms. It details how the model was evaluated and tested for risks, and what mitigations were implemented. Similarly, Meta’s LLaMA 3 model includes structured documentation outlining its architecture, training data, intended uses, and known limitations. These examples show that even the most powerful models benefit from clear, honest communication about their boundaries.
In the open-source ecosystem, Hugging Face has become a leader in promoting model cards. Developers uploading models to their platform are encouraged to fill out cards that explain their model’s performance, ethical considerations, and use-case suitability. Yet, as recent research shows, many model cards remain incomplete. One analysis of over 32,000 model cards found that only a small percentage addressed key issues like fairness, subgroup performance, or known risks (Gao et al., 2024). In practice, many cards simply recycle generic statements rather than documenting the real-world behavior of the specific model (Rahimi et al., 2024).
This inconsistency is problematic — especially in public health, where the stakes are high. When a predictive model is used to triage limited resources during a disease outbreak or to identify high-risk individuals for intervention, we need to know: Was this model validated across different racial and ethnic groups? How does it perform in low-data settings? What happens when the input data shifts due to a new health crisis? Without answers to these questions, public health leaders are left making decisions based on assumptions, not evidence.
The good news is that momentum around model cards is growing. Amazon, Google, and OpenAI are all integrating model documentation into their AI governance frameworks (AWS, 2023). New tools like AutoLLM-CARD are even exploring how to automatically extract documentation from research papers and code repositories to make model cards easier to generate. In healthcare and biomedical AI, researchers are proposing “layered” or even graphical model cards — visuals that clinicians can read quickly to understand a tool’s scope and constraints.
In public health education and professional training, model cards should be treated as core literacy — a skillset as essential as interpreting epidemiological data or evaluating clinical trial results. Schools of public health can lead the way by incorporating AI documentation review into coursework and capstone projects. Public health agencies can require model cards as part of procurement or evaluation processes when adopting AI tools. Even small steps — like requiring documentation before deployment — can help build a culture of responsible AI use.
Ultimately, model cards are not about slowing down innovation. They are about responsible acceleration — helping us move forward with confidence, clarity, and care. By demystifying the models we use, we can ensure that AI serves the values at the heart of public health: equity, trust, and the well-being of communities.
References
-
Hugging Face. (2024). Model cards documentation. https://huggingface.co/docs/hub/en/model-cards
-
OpenAI. (2023). GPT-4 System Card. https://cdn.openai.com/papers/gpt-4-system-card.pdf
-
Meta. (2024). LLaMA 3 MODEL_CARD.md. https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md
-
Gao, M., et al. (2024). Scaling Model Cards: A Study of 32,000+ Documentation Artifacts. https://arxiv.org/abs/2402.05160
-
Rahimi, F., et al. (2024). RiskRAG: A Data-Driven Solution for Improved AI Model Risk Reporting. https://arxiv.org/abs/2504.08952
-
Amazon Web Services. (2023). SageMaker Model Cards Documentation. https://docs.aws.amazon.com/sagemaker/latest/dg/model-cards.html
-
Hoeijmakers, R. (2024). Graphical Model Cards in Healthcare. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11861263
-
Zhang, K., et al. (2024). AutoLLM-CARD: Towards a Description and Landscape of Large Language Models. https://arxiv.org/abs/2409.17011
This article was generated with assistance from artificial intelligence. All sources have been verified as of the date of publication. (October 9,2025)