Model Card

A model card is a short, standardized document that travels with a trained machine-learning (ML) model and explains, in plain terms, what the model is, what it is for, how well it works, and where it should not be trusted. The idea was proposed in a 2019 research paper, Model Cards for Model Reporting, and it borrows deliberately from other fields: a nutrition label on food, a datasheet for an electronic component, a safety data sheet for a chemical. A typical model card covers model details (who built it, what version), intended use and out-of-scope uses, the factors and metrics it was evaluated on, the evaluation and training data, quantitative results — ideally broken down across different groups of people — and, importantly, ethical considerations, caveats, and recommendations. It is documentation aimed at whoever might use or be affected by the model.

The purpose of a model card is transparency and accountability. A trained model is otherwise a black box — you can run it, but you cannot easily tell what it was built for, how it performs on people unlike the average, or where it is likely to fail. A model card makes those things explicit, so a would-be user can judge whether the model fits their situation and where its outputs deserve caution. By reporting performance across subgroups, it surfaces fairness gaps that an overall accuracy number hides; by stating intended and out-of-scope uses, it discourages people from applying a model where it was never validated; by listing limitations and ethical risks, it invites responsible deployment rather than blind trust. In short, the card turns tacit knowledge trapped with the model's authors into something a reader can act on.

A model card documents a model; it is easy to confuse with documentation of the data that trained the model, and the two are complementary but distinct. A datasheet for a dataset (a related idea from the same research tradition) describes how a dataset was collected, what it contains, its biases, and its intended uses. A model card describes the trained model — its intended use, its measured performance, and its limitations — which of course depends on the data but is not the same thing. Reading only one leaves a gap: you might know a dataset's flaws without knowing how the resulting model behaves, or know a model's headline accuracy without knowing what data shaped it. Strong documentation practice pairs them, so a reader can trace behavior back to data and forward to appropriate use.

It is worth being honest about the state of model cards. As of the mid-2020s there is no legal requirement to publish one and no single mandated format or standard for their content, so cards vary widely in depth and candor. A thin card that lists a model's name and an overall accuracy figure, while omitting subgroup performance, out-of-scope uses, and ethical risks, meets the letter of the idea while missing its point — real transparency lives in the parts that are uncomfortable to write. A model card is only as useful as it is honest and specific. The best cards state clearly what the model should not be used for and where it is known to underperform, because those are exactly the things a careless deployment would otherwise discover the hard way.

If you are consuming a model — including an off-the-shelf model for marketing tasks like lead scoring, content generation, or audience prediction — read its model card before you deploy, and read the uncomfortable sections first. Check the intended-use and out-of-scope statements against your actual use; a model validated for one population or task may quietly fail on yours. Look for performance broken down by group, not just a single accuracy number, because a good average can hide poor results for the people you most need to serve well. Note the stated limitations and ethical risks and design guardrails around them. If a model has no card, or only a superficial one, treat that absence as a caution in itself.

If you are producing a model, write the card as a genuine account of the model's fitness, not a marketing sheet. Report subgroup performance even when it is unflattering, state the uses the model is not validated for, and name the ethical risks and failure modes plainly. The failures to avoid are the mirror image of good practice: a card that omits subgroup metrics and hides fairness gaps, one that overstates intended use so the model gets applied where it was never tested, one that skips limitations and ethical considerations, and one so vague it documents nothing. A model card done right is a small document that prevents large mistakes — it lets the people downstream of a model use it within its limits instead of discovering those limits in production.

The model card — standardized documentation of a machine-learning model's intended use, performance, limits, and ethical risks — was proposed in the 2019 paper Model Cards for Model Reporting to bring transparency to deployed models.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where model card is a core concern: