Model Cards
Description
A model card is a short structured document published with a trained model, stating what the model is for and what it should not be used for. It reports how the model performs on stated evaluation data, with results broken down by group (for example age or sex) and by operating condition (for example lighting or device) where these matter. It also records the training data, the metrics chosen and why, ethical considerations and known caveats. Readers use the card to judge whether the model suits their purpose before they deploy it. The original proposal gives suggested sections rather than a fixed format, and a card is only as reliable as the evaluation behind it.
Example Use Cases
Safety
Documenting a medical diagnosis AI with detailed performance metrics across different patient demographics, age groups, and clinical conditions, enabling healthcare providers to understand when the model should be trusted and when additional expert consultation is needed for patient safety.
Fairness
Publishing a model card for a CV-screening model that reports selection rates and error rates for each gender and age band. An HR team deciding whether to adopt the tool can see where results differ between groups and ask the vendor for further evidence before using it.
Transparency
Publishing detailed model documentation for a credit scoring API that clearly describes training data sources, evaluation methodologies, and performance limitations, enabling financial institutions to make informed decisions about model deployment and regulatory compliance.
General
A public body that buys AI tools asks each supplier for a model card and files it in its register of deployed models. Before approving a tool for a new use, officials compare the card's stated intended use and evaluation data with the proposed use and record any gaps.
Limitations
- Group-wise results need evaluation data labelled with the relevant attributes, which is often unavailable or restricted for privacy and legal reasons. Teams then report fewer groups than matter, and building the card takes time and statistical skill that can delay release.
- Cards are written by the developer and nothing in the format requires an independent check. Unflattering results or commercially sensitive details can be left out, and a study of 32,111 Hugging Face model cards found the evaluation and limitations sections among the least often completed.
- A card describes the model as evaluated at one point, on particular test data. Retraining, new deployment settings and data drift make it stale, and it stays in circulation unless someone owns updates and versioning.
- A machine-readable model card format now exists (the model card object in the CycloneDX bill of materials, defined in ECMA-424, second edition, December 2025). The original template is still a set of suggested sections, and cards differ widely in structure and depth, which makes them hard to compare across models and vendors.
- A card informs readers but does not constrain use or show that the model is safe. It cannot stop deployment outside the stated intended use, and the evidence it carries is only as good as the evaluations behind it.
Resources
Research Papers
Model Cards for Model Reporting
Trained machine learning models are increasingly used to perform high-impact tasks in areas such as law enforcement, medicine, education, and employment. In order to clarify the intended use cases of machine learning models and minimize their usage in contexts for which they are not well suited, we recommend that released models be accompanied by documentation detailing their performance characteristics. In this paper, we propose a framework that we call model cards, to encourage such transparent model reporting. Model cards are short documents accompanying trained machine learning models that provide benchmarked evaluation in a variety of conditions, such as across different cultural, demographic, or phenotypic groups (e.g., race, geographic location, sex, Fitzpatrick skin type) and intersectional groups (e.g., age and race, or sex and Fitzpatrick skin type) that are relevant to the intended application domains. Model cards also disclose the context in which models are intended to be used, details of the performance evaluation procedures, and other relevant information. While we focus primarily on human-centered machine learning models in the application fields of computer vision and natural language processing, this framework can be used to document any trained machine learning model. To solidify the concept, we provide cards for two supervised models: One trained to detect smiling faces in images, and one trained to detect toxic comments in text. We propose model cards as a step towards the responsible democratization of machine learning and related AI technology, increasing transparency into how well AI technology works. We hope this work encourages those releasing trained machine learning models to accompany model releases with similar detailed evaluation numbers and other relevant documentation.