Definition · Plain-language
Model card
A model card is a short, standardised document describing a machine learning model’s intended use, performance across conditions, limitations and ethical considerations.
The step most authors miss
Doing CRediT right? Don’t stop at the statement.
A CRediT statement credits you inside one paper. The recognition CRediT was built for happens when those roles are tied to you, persistently. Sign in with your ORCID — free — and claim your CRediT contributions on casrai.org, the home of the standard. They become a verified, portable part of your identity, not a line that disappears into one PDF.
Free: claim your contributions, then export a journal-ready CRediT statement, schema.org structured data, JATS XML, CSV or BibTeX — and preview your public profile. A membership publishes that profile publicly and verifies the journals you serve.
What a model card contains
A model card gathers, in one concise document, the information needed to judge whether a model suits a particular use. Typical sections cover the model’s intended use and out-of-scope uses; its training data and approach at a high level; quantitative performance, ideally broken down across relevant conditions and demographic groups rather than reported as a single headline figure; known limitations and failure modes; and ethical considerations such as fairness risks and recommended safeguards. The format proposed by Mitchell et al. in their 2019 paper "Model Cards for Model Reporting" deliberately keeps cards short and readable so they are actually produced and consulted.
Why model cards matter
Models are frequently reused in contexts their developers never tested, which is a common source of harm. A model card counters this by making intended use, validated conditions and limitations explicit, so a prospective deployer can see where a model should and should not be applied. Reporting performance disaggregated across groups also surfaces fairness problems that an aggregate accuracy figure would hide. Because they package exactly the information transparency and audit require, model cards have become a standard governance artefact and are widely adopted across model repositories and responsible-AI programmes as a lightweight but meaningful disclosure.
Model cards within governance
A model card is one of the documentation artefacts that operationalise transparency and accountability. It supports an AI audit by providing a baseline of claimed use and performance to test against; it supports risk management by recording limitations and ethical considerations; and it supports accountability by tying a model to its intended use and owner. Model cards are often paired with related artefacts such as datasheets for the underlying data and, in regulated settings, with the more formal technical documentation some regimes require. They are not a substitute for testing or oversight, but they are a key carrier of the information those activities depend on.
Key facts
At a glance
- Definition: a short standardised document describing an ML model’s use, performance, limits and ethics
- Origin: Mitchell et al., "Model Cards for Model Reporting" (2019)
- Key content: intended and out-of-scope use, disaggregated performance, limitations
- Purpose: help deployers judge if a model fits a context
- Role: a practical transparency and documentation artefact
- Related: datasheets for datasets; formal technical documentation
Common misconceptions
What people often get wrong
Often heard: A model card is just a marketing summary of a model.
Actually: A model card is a candid documentation artefact that records limitations, out-of-scope uses and performance across groups — including where a model performs poorly. Its value lies precisely in disclosing weaknesses, not in promotion.
Often heard: A single accuracy number is enough; disaggregation is optional.
Actually: A central recommendation of model cards is reporting performance broken down across relevant conditions and groups. An aggregate figure can hide serious disparities that only disaggregated reporting reveals.
Often heard: Producing a model card means a model is safe and approved.
Actually: A model card documents a model; it does not test, certify or approve it. It supports audit, oversight and informed deployment decisions but does not replace them.
Going deeper
