Practise vocabulary for fairness metrics, bias audits, EU AI Act risk tiers, and communicating ML limitations to stakeholders.
0 / 6 completed
1 / 6
Demographic parity as a fairness metric requires:
Demographic parity (statistical parity) requires equal positive prediction rates across groups. This is one of several competing fairness definitions — satisfying one often makes others impossible.
2 / 6
When writing a bias audit summary for stakeholders, the most important element is:
Stakeholders need to understand the business and legal implications of bias, not the technical details. A good bias audit summary explains what disparity was found, which user groups are affected, what the risk is, and what will be done about it.
3 / 6
The EU AI Act classifies AI systems as 'high-risk' when:
EU AI Act high-risk categories include: biometric identification, critical infrastructure, education, employment, essential services (credit, insurance), law enforcement, migration, and justice. High-risk systems face strict requirements including conformity assessments and transparency obligations.
4 / 6
A model exhibits 'representational harm' when:
Representational harms are distinct from allocative harms (unfair distribution of resources). Examples: a translation model defaulting to male pronouns for doctors, or an image model generating stereotyped representations of certain ethnicities.
5 / 6
The correct way to communicate model uncertainty to a business stakeholder is:
Stakeholders need to make decisions based on model outputs. Translating uncertainty into business-relevant terms ('the model is confident for high-value transactions but less reliable for first-time customers') enables better risk-aware decisions than statistical jargon.
6 / 6
Model cards are used to:
Model cards (Mitchell et al., 2019) are transparency documents covering: model details, intended use, out-of-scope uses, factors affecting performance, evaluation metrics across subgroups, ethical considerations, and caveats.