LANDMARK · 6 MIN · UNDERSTAND
Model cards and documentation
After this landmark, you can extract the decision-relevant facts from a model card, and recognise when its silences are the important part.
↗
This concept is shared, but the Everyday lens is less central here.
You can still explore it. We’re showing the shared explanation and a related practical view without hiding the knowledge.
A model card is a short structured document published alongside a model: what it is, what it was intended for, how it was evaluated, where it performs poorly, and what uses are out of scope. Good ones also state the training-data sources at least in outline, the evaluation results broken down by group rather than only in aggregate, the safety testing performed, and the known limitations in plain language. Model cards began as a research-community practice and are becoming a regulatory expectation, because transparency obligations require exactly this kind of documentation to exist. For a reader, the skill is knowing which parts change a decision. Intended use and out-of-scope use tell you whether your application is one the developer tested for. Disaggregated evaluation tells you whether good average performance hides a badly-served group. And an absent section is information: a card that reports only aggregate scores and no limitations has told you the developer either did not measure or chose not to say.
This is the canonical concept. It stays the same across learner lenses so personalization never changes the underlying facts.
◌ EVERYDAY LENS
What this looks like for you
You will rarely read one, but it’s worth knowing they exist. When coverage says a model is “tested for safety”, there is usually a document stating what was actually tested — and it is often narrower than the headline implies.
MAKE A DECISION
A model card reports strong aggregate accuracy on speech transcription but gives no breakdown by accent or speaker. You need it for a public-facing service. What follows?
CARRY THISOpen the model card for a model your organisation uses. Note one thing it says about out-of-scope use, and one thing it does not report at all.