Module 2 self-check | AI GRCA Practitioner Programme

When should a harm block a launch?

Never — harms are always accepted with monitoringWhen it is severe and/or probable enough that available controls can't reduce it to an acceptable levelOnly if a regulator has already fined someone for itOnly for generative AI systems

A model quietly moving from 'decision support' to being treated as an automatic decision-maker illustrates a bias/harm risk at which lifecycle stage?

Model architecture selectionNone — this isn't a riskData and labelsDeployment and use

Which best distinguishes accuracy from fairness?

They mean the same thingHigh overall accuracy can still hide unequal error rates across groupsFairness only matters for generative AIAccuracy always implies fairness

The 'green apple bias' example is typically used to illustrate:

A copyright issueA model performing worse on cases that look different from its dominant training examplesA pricing algorithm errorA cybersecurity vulnerability

Why might 'monitor' be an acceptable treatment for a low-probability, low-severity harm, when it wouldn't be for a high-severity one?

Monitoring replaces the need to ever assess severityMonitoring is never an acceptable treatment for any harmTreatment should be proportionate to the risk level; lower risks can sometimes be managed with monitoring aloneAll harms require the same treatment regardless of severity

What is the relationship between transparency and explainability?

Transparency only concerns pricingThey are identical conceptsTransparency is about making information about a system available; explainability is about understanding why a specific output occurredExplainability only applies to generative AI

An 84% 'overall accuracy' figure for a learner engagement score says what about fairness across student groups?

Accuracy and fairness are unrelated concepts entirelyIt proves the system is fair to all groupsNothing on its own — subgroup performance must be separately testedIt proves the system is unfair

An umbrella (protection against a likely, low-severity event — rain) is often used to illustrate which risk concept?

That treatment should match the probability and severity of the specific risk, not be one-size-fits-allThat all risks should always be acceptedData minimisationRisk transfer only

Responsible AI, Trustworthy AI & Harm