AI Governance and Ethics practice questions

Pseudonymisation by replacing patient names with random IDs

Why it fails: Replacing names with random IDs leaves other identifiers such as dates of birth, postcodes, and rare diagnoses intact, so records remain re-identifiable; HIPAA's de-identification standard requires removal or transformation of all such identifiers. Pseudonymisation suits internal linkage and audit trails, not releasing data for model training.

Data augmentation to create synthetic samples

Why it fails: Augmentation generates additional synthetic records from existing ones, which does not remove identifiers or break the link between a record and the individual, so re-identification risk is unchanged. It tempts because synthetic data can reduce exposure, but that requires fully generative replacement, not augmentation of real patient records.

Data minimisation by removing all features except age and gender

Why it fails: Stripping all features except age and gender destroys the clinical signal needed to predict disease risk, so the model cannot meet its purpose; HIPAA permits de-identification while retaining relevant attributes. It tempts because minimisation is a genuine privacy principle, but it applies to collecting only necessary fields, not discarding diagnostic data.

Use more GPUs to parallelise training and reduce wall-clock time

Why it fails: Adding GPUs raises total energy consumption even when wall-clock time falls, so it does not align with minimising environmental impact. It is tempting because parallelism shortens training duration, but green AI weighs total energy and carbon, favouring efficiency, smaller models and renewable-powered hardware.

Deploy the model on a cloud provider with renewable energy certificates

Why it fails: Renewable energy certificates offset consumption on paper but do not reduce the compute energy a from-scratch training run actually draws, so emissions from GPU hours remain. Certificates suit organisations reporting purchased-energy totals across steady workloads, not cutting the training footprint itself.

Train the model on a larger dataset to improve accuracy

Why it fails: Training on a larger dataset increases compute hours and energy consumption, directly raising the carbon footprint of model training. Larger datasets serve accuracy gains, not environmental efficiency. This would suit a scenario prioritising model performance over sustainability, where additional training cost is acceptable.

Demographic parity violation

Why it fails: Demographic parity concerns the proportion of positive outcomes, not the calibration of probabilities.

Equalised odds violation

Why it fails: Equalised odds requires equal true positive and false positive rates across groups, but the scenario describes a systematic probability shift without a corresponding difference in actual outcomes—this violates calibration, not equalised odds. It is tempting because equalised odds also addresses group-level fairness, and would be correct if the model’s decision threshold produced disparate error rates for the same outcome likelihood.

Individual fairness violation

Why it fails: Individual fairness requires similar predictions for similar individuals, not calibration across groups.

Use FP32 precision instead of mixed precision

Why it fails: FP32 requires more memory and compute than mixed precision (FP16), increasing energy use.

Use more GPUs in parallel to finish training faster

Why it fails: More GPUs increase total energy consumption even if wall-clock time decreases, unless the workload perfectly scales.

Increase the number of training epochs for better accuracy

Why it fails: More epochs mean more computation and thus higher energy consumption.

Implement human-on-the-loop oversight where humans only monitor outputs

Why it fails: Human-on-the-loop monitoring alone leaves no authority to intervene before a consequential teacher evaluation is issued, so contested or biased outputs stand. It is tempting because it scales cheaply for low-stakes, high-volume tasks, but high-stakes decisions need human-in-the-loop review with power to override.

Publish the entire training dataset publicly for transparency

Why it fails: Publishing the full training dataset exposes personally identifiable pupil and teacher records, breaching data protection without proving model fairness. It is tempting because transparency builds trust, but that is the correct approach for openly licensed datasets, not for sensitive personal data requiring anonymisation and controlled access.

Require the system to be built only with open-source models

Why it fails: Open-source licensing says nothing about fairness, validity or accountability for high-stakes teacher evaluation; a proprietary model with strong bias testing and audit trails can be responsible. It is tempting because open weights aid scrutiny, but that is the correct requirement when transparency of model internals is the governance objective.

Train on larger datasets to improve accuracy

Why it fails: Larger datasets require more compute and energy, increasing the carbon footprint.

Use more powerful GPUs to speed up training

Why it fails: More powerful GPUs may speed up training but often consume more energy, increasing the carbon footprint.

Use older, less efficient hardware to save on manufacturing emissions

Why it fails: Older hardware is typically less energy-efficient, leading to higher operational emissions.

Offer the patient the option to have the decision reviewed by a human

Why it fails: Human review is a separate right (automated decision-making), but does not by itself satisfy the right to an explanation of the original decision.

State that the decision was made by an AI system and cannot be explained

Why it fails: This does not satisfy the right to explanation; GDPR requires that the data subject be given meaningful information about the logic.

Provide a copy of the model's source code to the patient

Why it fails: Source code is not understandable to most patients and does not explain the logic of a specific decision.

Limited risk

Why it fails: Limited risk applies to systems with specific transparency obligations, such as chatbots.

Minimal risk

Why it fails: Minimal risk is a residual category for all other AI systems.

General purpose AI

Why it fails: General purpose AI is a separate category not directly tied to risk tiers.

Unacceptable risk

Why it fails: Unacceptable risk includes systems like social scoring by public authorities, not credit scoring.

Obtaining a CE marking of conformity through self-assessment only

Why it fails: High-risk systems typically require third-party conformity assessment unless a harmonised standard covers all aspects.

Acquiring liability insurance for AI-related harms

Why it fails: The EU AI Act imposes no mandatory liability insurance requirement for high-risk systems; obligations centre on risk management, data governance, technical documentation, logging, and human oversight. Insurance is a commercial risk-transfer choice. It is tempting because financial institutions routinely insure AI risks, but this is not a regulatory precondition for market placement.

Publishing the training dataset in an open repository

Why it fails: Publishing training datasets openly is not mandated; the Act requires data governance, documentation, and registration in an EU database, not public disclosure. It is tempting because transparency duties exist, but they concern information provision to deployers and authorities, not open-sourcing proprietary training data.

Publish the model's source code publicly

Why it fails: The Act does not require open-sourcing of high-risk AI systems.

Conduct a data protection impact assessment (DPIA)

Why it fails: DPIA is required under GDPR, but not specifically mandated by the EU AI Act for all high-risk systems.

Disclose to users that they are interacting with an AI system

Why it fails: Transparency obligations apply, but not uniquely to high-risk systems; also, disclosure is not always mandatory for back-end processing.

Differential privacy

Why it fails: Differential privacy is a mathematical privacy guarantee but not a HIPAA-specific requirement.

Anonymisation

Why it fails: Anonymisation is broader; HIPAA's Safe Harbor specifically uses de-identification.

Pseudonymisation

Why it fails: Pseudonymisation is reversible and does not fully comply with HIPAA's de-identification standard.

Human-in-command approach

Why it fails: Human-in-command is a broader concept; HITL is the specific mechanism mandated.

Human-on-the-loop oversight

Why it fails: Human-on-the-loop is a form of oversight but not the specific requirement; HITL is more explicit for high-risk systems.

Automated decision-making without human review

Why it fails: High-risk systems require human oversight, so fully automated decisions without review are prohibited.

Govern - establish governance policies

Why it fails: Govern is about policies, but Map comes first to understand the risk context.

Manage - implement mitigation strategies

Why it fails: Management happens after measurement.

Measure - evaluate the fairness metrics

Why it fails: Measurement occurs after mapping the context.

Definition of roles and responsibilities for AI development

Why it fails: While important, roles are often defined separately from the policy.

Vendor AI assessment procedures

Why it fails: Vendor assessment is part of procurement, not the internal AI policy.

Data minimization requirements

Why it fails: Data minimization is a privacy principle, not a core policy component.

Attention visualization

Why it fails: Attention mechanisms are specific to transformer models and may not apply to other architectures.

Model cards

Why it fails: Model cards document overall model behavior, not individual predictions.

SHAP values

Why it fails: SHAP provides global and local feature importance based on game theory, but LIME is more commonly used for local explanations.

Individual fairness

Why it fails: Requires similar individuals to be treated similarly, but harder to operationalize in this context.

Accuracy parity

Why it fails: Not a standard fairness metric; accuracy should be consistent but is not a specific fairness metric.

Calibration

Why it fails: Ensures predicted probabilities match actual outcomes, but does not address group fairness directly.

The system is prohibited because it uses AI to infer emotions in a workplace or educational setting, but retail stores are exempt.

Why it fails: The EU AI Act prohibits emotion inference in workplace and educational settings, but it also prohibits emotion recognition in other contexts if it uses subliminal techniques or exploits vulnerabilities. However, this system operates in retail, which is not explicitly exempt. The prohibition on emotion inference in retail is not absolute; it depends on the technique and purpose.

The system is high-risk and must undergo a conformity assessment before deployment.

Why it fails: While emotion recognition systems can be high-risk in some contexts, the EU AI Act specifically prohibits real-time emotion inference in retail settings when used for commercial purposes. High-risk classification applies to other emotion recognition uses, such as in recruitment or education. Here, the real-time commercial use triggers a prohibition, not a high-risk classification.

The system is allowed as long as customers provide explicit consent and the company publishes a transparency notice.

Why it fails: Consent and transparency do not override prohibitions under the EU AI Act. Prohibited practices are banned outright, regardless of consent. The Act bans emotion inference in commercial contexts to prevent manipulative practices. Thus, even with consent, this system cannot be deployed in the EU.

Calibration

Why it fails: Calibration ensures predicted probabilities match observed outcome frequencies within groups; it does not compare positive prediction rates. It is tempting because it is a group-aware fairness measure, and would be correct when the requirement is that predicted risk scores are equally reliable across groups, not equal approval rates.

Individual fairness

Why it fails: Individual fairness requires similar predictions for similar individuals, not equal positive prediction rates across groups. It is tempting because it addresses discrimination at the person level, and would be correct when the requirement is consistency between comparable individuals rather than group-level parity of approval rates.

Equalised odds

Why it fails: Equalised odds requires equal true positive and false positive rates across groups, not equal positive prediction rates. It is tempting because it is a group fairness metric, and would be correct when the requirement is parity of error rates conditional on the actual outcome, rather than demographic parity of approvals.

Full transparency by publishing the model's source code and training data

Why it fails: Publishing source code and training data is not an EU AI Act requirement for high-risk systems; that obligation concerns GPAI models, not employment screening. High-risk providers must instead implement risk management, technical documentation, logging, human oversight and conformity assessment. Transparency here means informing affected persons, not open-sourcing proprietary models.

Annual third-party audits of the model's energy consumption

Why it fails: Energy-consumption audits address environmental impact, not the EU AI Act's high-risk requirements of risk management, data governance, technical documentation, logging, transparency, human oversight and conformity assessment. It is tempting because sustainability auditing is topical, but it does not satisfy the Act's high-risk obligations.

Obtaining explicit consent from each applicant to process their data

Why it fails: Explicit consent is a GDPR data-processing condition, not an EU AI Act high-risk obligation; the Act instead requires risk management, technical documentation, logging, human oversight and conformity assessment. It is tempting because consent feels protective, but consent alone does not satisfy the high-risk regime.

A SHAP summary plot showing mean absolute SHAP values across all features

Why it fails: A summary plot aggregates mean absolute SHAP values across the whole dataset, describing global feature importance rather than one prediction. It would be the right choice when explaining overall model behaviour to stakeholders. For a single prediction, a local force or waterfall plot showing that instance's feature contributions is needed.

A SHAP dependence plot for the top feature

Why it fails: A dependence plot shows how one feature's SHAP value varies with its value across many instances, revealing global relationships and interactions. It answers how a feature influences predictions generally. Explaining a single prediction needs a local waterfall or force plot covering all contributing features for that instance.

A SHAP bar chart of absolute feature importance

Why it fails: A bar chart of absolute feature importance again aggregates magnitude across all instances, giving a global ranking with no direction or per-instance detail. It suits reporting which features matter overall. Explaining one prediction requires a local plot, such as a waterfall, that decomposes that instance's SHAP values.