What Is a Model Card? The Nutrition Label for AI Models
Every serious open-weights release ships with a document that almost nobody reads in full: the model card. Frontier labs publish longer cousins called system cards, sometimes running past 150 pages. These documents sit at the intersection of engineering practice, platform norms, and now law. Here is what they are, where they came from, and what to actually look for in one.
The idea: document the model like you document a component
A model card is a structured description that accompanies a machine learning model. It answers a small set of standing questions: what the model does, what it was trained on, how it was evaluated, where it performs well, where it fails, and what it should not be used for.
The concept comes from a 2019 paper, "Model Cards for Model Reporting," by Margaret Mitchell, Timnit Gebru and seven co-authors at Google (FAccT 2019). The authors borrowed from two mature documentation habits: component datasheets in electronics, and the information sheets that come with prescription drugs. The paper's running example was a card for the Perspective API toxicity classifier, written after researchers found the model produced oddly high toxicity scores for certain benign sentences containing identity words like "gay". The card existed to stop exactly that kind of surprise: a team adopting a model without knowing its failure modes.
The same research circle produced the companion idea for data. Gebru and Hannah Wallach's "Datasheets for Datasets" (2018) argued that datasets, not just models, need provenance documentation. Model cards and datasheets were designed as halves of one system: one documents the artifact, the other documents the raw material.
Model card, system card, same job
The terminology has drifted. In current practice:
- A model card documents the model itself: architecture, training data, evaluations, intended and out-of-scope uses, known limitations. Google DeepMind publishes a model card for every Gemini variant, updated over time, listed on a dedicated page.
- A system card covers the model as deployed, with more weight on safety evaluation: red-team results, misuse probes, jailbreak resistance, capability thresholds and the deployment decisions they trigger. OpenAI introduced the term in the GPT-4 technical report of March 2023, which devoted a large dedicated section to a system card. Anthropic now publishes a system card for essentially every Claude release, with dated entries going back to Claude 2 in 2023.
The line between the two is a convention, not a standard. There is no mandatory schema for either.
The recent Claude system cards show how far the genre has run from a label. The Claude Opus 4.7 card exceeds 200 pages and includes dedicated chapters on cyber capability evaluations, reward-hacking tests and autonomy metrics. The Claude 4 card from May 2025 drew wide attention because its agentic misalignment evaluations described scenarios in which the model threatened blackmail or leaking in fictional corporate-hostage setups. Whatever else they do, system cards have become a primary channel through which labs disclose how their models behave under pressure, and a favorite quarry for researchers picking those disclosures apart.

The platform made cards the default
Hugging Face turned the model card from a proposal into infrastructure. On the Hub, every model repo's README.md is rendered as a model card: a markdown file with a YAML metadata block at the top recording base models, licenses, task tags and evaluation results. The Hub documentation calls cards "essential for discoverability, reproducibility, and sharing," and the company ships a card template, an annotated guide to filling it out, and a creation tool.
Adoption is broad. Completeness is not. A February 2024 study analyzed 32,111 model documentations on Hugging Face. Most models with substantial downloads do provide a card, but informativeness is uneven: the training section is the most consistently filled out, while the sections users most want are the emptiest. Environmental impact appeared in roughly 2% of cards, evaluation details in about 15%. The researchers also ran an intervention, writing proper cards for 42 popular models that had sparse or no documentation, and found the added cards were moderately correlated with higher weekly downloads. Documentation, in other words, has a measurable commercial side effect.
A separate 2025 audit of Hub metadata found documentation "often incomplete, malformed, or available only in human-readable text" even for large projects. One example it gives: OpenAI's clip-vit-large-patch14-336, with millions of downloads, carries a card that is almost entirely empty.

It is becoming a legal requirement
The EU AI Act converted the norm into an obligation for at least some of this documentation. Since August 2, 2025, providers of general-purpose AI models placed on the EU market must:
- draw up and maintain technical documentation covering training, testing and evaluation results, with a minimum content list in Annex XI (tasks the model performs, acceptable use policies, architecture and parameter count, training data provenance and curation, compute used, estimated energy consumption);
- provide information to downstream companies that integrate the model, per Annex XII;
- publish a summary of training content used for copyright compliance purposes.
The documentation itself goes to the AI Office and national authorities on request; the training-data summary is public. Open-source models are exempt from parts of this unless they are classified as systemic-risk, and models already on the market before August 2025 have until August 2027 to comply.
Annex XI reads like a model card with legal force behind it. That is not a coincidence: the Annex's fields for provenance, curation and evaluation mirror what the documentation literature has proposed since 2018.
How to read a card without wasting an afternoon
Cards are self-reported, unreviewed prose. Two documents with identical section headers can differ by three orders of magnitude in honesty. A practical reading order:
- Limitations and intended-use sections. The most informative part of a good card, and the first thing to check is whether it is specific. "Not suitable for medical advice" is boilerplate; a list of measured failure distributions is disclosure.
- Evaluation methodology, not the headline number. A benchmark score without task configuration, prompting format or evaluation harness details is marketing. This matches what the 32K study found in reverse: evaluation is the section practitioners most often skip.
- Training data provenance. Provenance statements are usually vague by design, but what a lab will say at all tells you about exposure, cutoff and curation.
- License and base model lineage. Fine-tunes inherit the base model's license and limitations, and a card that names no base model deserves suspicion.
- For system cards, the safety appendix, not the press release. The interesting claims (what was tested, what thresholds triggered what action) live in the fine print, and independent analysts routinely disagree with the labs' own summaries of it.
The genre's flaw was visible in its founding example: the Perspective API surprise was discovered by users doing their own evaluations, years after the classifier shipped widely. Cards are best read as a floor for due diligence, not a substitute for it.
Sources
- Mitchell et al., "Model Cards for Model Reporting," FAccT 2019: https://arxiv.org/abs/1810.03993
- Hugging Face, Model Cards documentation: https://huggingface.co/docs/hub/en/model-cards
- Hugging Face blog, "Model Cards": https://github.com/huggingface/blog/blob/main/model-cards.md
- Liang et al., "What's documented in AI? Systematic Analysis of 32K AI Model Cards" (Feb 2024): https://arxiv.org/abs/2402.05160
- "An Empirical Analysis of Machine Learning Model and Dataset Documentation, Supply Chain, and Licensing Challenges on Hugging Face" (2025): https://arxiv.org/abs/2502.04484
- Anthropic, Model system cards index: https://www.anthropic.com/system-cards
- Anthropic, Claude Opus 4.7 System Card (April 2026): https://www-cdn.anthropic.com/037f06850df7fbe871e206dad004c3db5fd50340/Claude%20Opus%204.7%20System%20Card.pdf
- Google DeepMind, Model cards page: https://deepmind.google/models/model-cards
- EU AI Act, Article 53 (Obligations for providers of general-purpose AI models): https://artificialintelligenceact.eu/article/53/
- EU AI Act, Annex XI (Technical documentation): https://artificialintelligenceact.eu/annex/11/
- Promptfoo, "System Cards Go Hard" (history of system card releases): https://www.promptfoo.dev/blog/system-cards-go-hard
