AIF-C01 - Responsible and secure AI (12% of the exam) - Section 5.1
Explain the development of responsible AI systems.
Recognise bias, fairness, and safety concerns, and the artefacts that address them, such as model cards that disclose intended use, limitations, training-data characteristics, and evaluation results.
Model cardsBias and fairnessResponsible AI
Practice question for this objective
lock_openFree sampleResponsible and secure AIeasy
A team is publishing a model card to accompany an image-classification model so that downstream adopters can judge whether it suits their own setting. They want the card to disclose the kind of information a model card is specifically intended to communicate. Which two pieces of information genuinely belong in the model card? Select TWO.
AThe cloud invoice and monthly hosting spend the team incurred while training the model on its accelerators.
BThe characteristics of the training data, including its sources, time range, and how representative it is of real conditions.check_circle Correct
CThe login credentials and access keys for the private repository where the model weights are stored.
DA marketing tagline asserting that the model is the most advanced classifier available on the market today.
EThe model's known limitations and evaluation results, including where its accuracy is measured to degrade.check_circle Correct
Recognise that a model card discloses training-data characteristics, intended use, limitations, and evaluation results so adopters can judge fitness before use. A model card is a transparency artefact written for downstream users. It describes what the model was trained on, how it was evaluated, and where its performance is known to break down, so adopters can decide whether it fits their context. Hosting costs, repository secrets, and marketing claims are not transparency content and do not belong in the card.
Why A is wrong: Tempting as project detail, but billing figures are an internal cost matter and are not part of what a model card discloses to adopters.
Why B is correct: Training-data characteristics are a core model card section so adopters can judge whether the data resembles their own population.
Why C is wrong: This looks like provenance, but secrets must never be published and are a security matter rather than transparency content.
Why D is wrong: A slogan is promotional filler, whereas a model card is meant to give factual, checkable disclosures rather than sales claims.
Why E is correct: Documenting limitations and evaluation outcomes lets adopters see where the model is reliable and where it fails before they use it.
Answers that look right on this material and are not. Each one is a distractor from a different question in the AIF-C01 bank for this domain.
A model card is a runtime load balancer that distributes inference traffic across replicas, while a model registry is the managed GPU cluster that schedules and executes the training jobs.
errorWhy it is wrong: Load balancers and training clusters are infrastructure components, not governance artefacts, so this answer swaps the documentation and cataloguing concepts for unrelated compute resources.
Measuring the exact monthly inference cost the model will incur once it is deployed at production traffic levels
errorWhy it is wrong: Cost depends on pricing and request volume rather than model documentation, so a model card does not provide the figures needed for a cost projection even though budgeting is a real adoption concern.
It records the live infrastructure metrics of the serving endpoint, such as latency and request volume, so operators can monitor production health.
errorWhy it is wrong: Endpoint latency and request volume are operational monitoring concerns handled by observability tooling, not the content of a model card, which documents the model itself rather than its runtime infrastructure.