Model Card
A standardised document describing what an AI model can do, how it was built, and where it falls short.
A Model Card (introduced by Google researchers Mitchell et al., 2019) is a short document that accompanies an AI model release, disclosing its intended use cases, training data characteristics, evaluation results, limitations, and ethical considerations.
Typical sections:
- Model description: Architecture, size, training objective
- Intended use: What tasks the model was designed for
- Out-of-scope use: What it shouldn't be used for
- Training data: Sources, size, known biases
- Evaluation results: Benchmark scores across demographic groups and tasks
- Limitations: Known failure modes, biases, capability gaps
- Ethical considerations: Potential harms, recommended safeguards
Why they matter: Model cards are the primary transparency mechanism for AI releases. Without them, deployers have no documented basis for assessing fitness for a given use case. Regulators in the EU (AI Act) and several US states now require disclosure documents that extend the model card concept.
Where to find them: Hugging Face model pages, Anthropic's model documentation, Google's model explorer, and academic papers accompanying open-source releases.
In plain terms
A nutrition label for AI: tells you what's inside, how much, and what you should and shouldn't use it for — so you can make an informed choice.
Related concepts
AI Safety
The field working to ensure AI systems do what humans actually want — now and as they become more capable.
AI Benchmark
Standardised tests that measure what a model can actually do.
Large Language Model
AI trained on vast text to understand and generate language.
Fine-tuning
Specialising a general AI model on your specific data and style.