What A Model Card Contains
A model card is the release document that travels with an AI model. It states what the model is for, how it performs across different segments and where it is known to fail. The AIGP exam treats the model card as part of release readiness. Understanding the document therefore beats memorising the term.
Where the model card came from
The idea has a specific origin. In 2019, Margaret Mitchell and colleagues published Model Cards for Model Reporting. The paper proposes short standardised documents to accompany trained machine learning models. Its framework runs to nine sections: model details, intended use, factors, metrics, evaluation data, training data, quantitative analyses, ethical considerations, plus caveats and recommendations.
The motivating problem was misuse through ignorance. A model trained and evaluated on one population gets deployed on another, and nobody documented the difference. Disaggregated performance reporting is the paper’s central demand, meaning results broken down by group rather than averaged.
What a model card is for
Three audiences read the card, and each takes something different from it.
The deployer
An organisation adopting a model needs to configure it responsibly. The card tells the deployer which use cases the developer intended, which it tested and which it warns against. A card that states inappropriate uses alongside appropriate ones is doing its job. Choosing how to run the model in the first place is a separate decision. That choice is worked through in Choosing Between AI Deployment Options.
The user and the affected person
People relying on outputs need to calibrate that reliance. Segment-level performance figures tell a user where the model is strong and where its error rate climbs. Averaged headline accuracy hides exactly the information this reader needs.
The governance function
For the AI governance professional, the card is release evidence. It shows that someone defined the purpose, measured performance against it and recorded the limits. A release decision can then attach conditions on the strength of what the card discloses, such as limited scope or enhanced monitoring.
The model card and the technical file
The EU AI Act never mandates a model card by that name. For high-risk systems it mandates technical documentation instead, under Article 11 of Regulation 2024/1689. That file is drawn up before the system reaches the market and kept up to date. Annex IV sets its minimum content. Its audience is national authorities and notified bodies, in enough depth to assess conformity.
The two documents differ in audience and depth. A technical file is a maintained compliance artefact for regulators. A model card is a readable release summary for deployers and users. Exam questions test precisely this confusion, because the wrong answers sit close together.
General-purpose models add a third document to the stack: the public summary of training content that the AI Act requires from their providers. That summary describes what went into training. Its ground connects to the copyright reservations covered in Text And Data Mining Opt-Outs.
How the AIGP exam tests the model card
The Body of Knowledge is the IAPP’s published outline of exam coverage. For AIGP, the domain on governing AI development places the model card inside release readiness. Assessing whether a system is ready for production includes creating the card and satisfying conformity requirements. Definition questions ask what the card is. Keep it distinct from the technical file, from public capability statements and from training-content summaries. The wrong answers are always its neighbours.
Application questions go one step further. They describe a deployment that ignored a stated limitation and ask where governance failed. The answer usually traces back to the card nobody read.
Before drilling questions, it helps to know which domains need the work. The free AIGP Assessment locates the gaps. For the question method itself, the AIGP Exam Question Masterclass at 195 euros takes apart exactly this kind of definition-and-neighbour item.
