Model card: A model card is a standardized document published alongside an AI model that describes what it was built to do, how it was trained and tested, where it performs well or poorly, and what its known risks and limitations are.
Think of a model card as the nutrition label or spec sheet for an AI model. The format was proposed in a 2019 research paper by Margaret Mitchell and colleagues as a way to make model makers state, in one place, the intended uses, the evaluation results broken out by group, and the situations where the model should not be used. Most major labs now publish something along these lines for each release, sometimes under the name 'system card' when it covers the whole product rather than the bare model.
A good card tells you the model's knowledge cutoff, its context window, the languages it handles, the benchmarks it was measured on, what safety testing was done, and which uses the vendor prohibits. It should also be honest about weaknesses: known biases, tasks it fails, and behaviors that changed from the previous version. Cards vary a lot in quality. Some are detailed technical reports; others are a marketing page with a table.
For buyers, the card is where due diligence starts. If a vendor cannot point you to documentation like this for the model under their product, that is itself useful information. Model cards are also becoming a compliance artifact, referenced by procurement checklists and by emerging AI regulation that asks providers to document their systems.
Example at work
A compliance officer reviewing a proposed AI contract-review tool asks the vendor which model powers it and pulls up that model's card. The card lists the knowledge cutoff, the safety evaluations run, and a note that the model is not intended for legal advice without human review. Those three facts go straight into the risk assessment and the usage policy.
Why it matters
A model card is the fastest way to learn what a model is and is not built for, straight from the people who built it. Reading one before adopting a tool takes twenty minutes and can save you from deploying the wrong model for a regulated or high-stakes task.
Related terms
- Responsible AIResponsible AI is an approach to building and using AI that deliberately addresses fairness, transparency, privacy, safety, and accountability, so the technology's benefits do not come at the expense of the people affected by it.
- AI governanceAI governance is the set of policies, roles, controls, and oversight processes an organization uses to decide how AI is adopted, used, monitored, and held accountable across the business.
- BenchmarkA benchmark is a standardized test used to measure and compare AI models on a specific skill, such as math, coding, knowledge, or reasoning, by scoring their answers against a fixed set of questions.
- Knowledge cutoffA knowledge cutoff is the date after which a language model's training data stops. Unless it is given live information through search or documents, the model knows nothing about events, prices, or rules that changed after that date.
- ExplainabilityExplainability is the degree to which you can understand why an AI system produced a particular output. It matters most when a decision affects a person and someone has to justify it.