A team is asked to produce documentation for a model that scores loan applications. They propose a technical write-up of the architecture and the training data. What is missing, who actually reads this document, and what is it not?

Who reads it

Three audiences, with different questions, and a document written only for the first fails:

• A reviewer or validator asks whether the model does what it claims and where it stops being reliable.
• A deployer — often a different team, sometimes a different company — asks what they are allowed to use it for.
• A regulator or auditor asks whether the obligations attaching to this kind of system have been met.

What the proposal is missing

The model card format published by Mitchell and colleagues in 2019 named the sections that technical write-ups usually omit, and they are the ones the last two audiences need:

• Intended use, and explicitly out-of-scope use. A model validated for one population and applied to another is the most common real failure, and the document is where that boundary is set.
• Performance disaggregated by relevant group, not a single aggregate number. An overall AUC hides a subgroup where the model is useless.
• Training data provenance and known limitations, including what is not represented.
• Ethical considerations and the evaluation done, including what was tried and failed.
• Version, date, owner and change history, because a document without them describes a model that may no longer be deployed.

For a consequential use such as credit, obligations attach: the EU AI Act, in force since 1 August 2024 with most high-risk obligations applying from 2 August 2026, requires technical documentation, risk management and human oversight for systems in its high-risk tier.

What it is not

It is not a substitute for monitoring, and this is the mistake to name. A document describes the model at a moment. The population drifts, the upstream features change, and the document stays accurate about a model that no longer behaves as described.

It is also not a place to argue the model is fine. A model card with no limitations section is a marketing document, and a reviewer reads its absence as a finding.

When not to write the heavyweight version

For an internal model whose output a human uses as one input among several, with no external party and no consequential decision, a one-page card is proportionate: intended use, data, headline performance, known weaknesses, owner. The heavyweight form is driven by consequence and by who must be convinced, not by the model's technical sophistication.