CDEF common mistakes
CDEF · Understand the traps and how to avoid them.
Answering ethical-theory questions by picking whichever option sounds most benevolent, without identifying which normative framework (consequentialist, deontological, virtue, care, contractualist) the scenario is testing.
Candidates memorise framework labels but never practise mapping a scenario's structure (rule vs. outcome vs. character vs. relationship) onto the label, so they default to 'the kindest-sounding answer'. Many distractors are deliberately written to sound compassionate while violating a duty or a right.
What to do instead
For every practice scenario, write one line before choosing: 'This asks about the locus of moral judgement: act/outcome, rule/duty, agent/character, or relation/care.' Then eliminate options whose locus does not match the stem. Drill with paired scenarios that are identical except for one structural feature so the framework shift becomes visible.
Treating 'consent' as a single yes/no checkbox and ignoring whether it was informed, specific, freely given, and revocable in the scenario described.
Consent appears in privacy, research-ethics, and governance domains, and candidates carry a colloquial sense of 'they agreed, so it's fine'. Exam items often describe consent obtained under employer pressure, bundled into unrelated terms, or given without understanding of downstream uses.
What to do instead
Apply a four-part test to every consent fact pattern: capacity, information (purpose, recipients, consequences), voluntariness (power imbalance, alternatives), and granularity/withdrawal. State explicitly which element fails. Practise spotting implicit coercion markers such as 'as a condition of employment' or 'standard terms accepted at sign-up'.
Confusing legal compliance with ethical adequacy, and selecting 'it is lawful, therefore it is acceptable' as the resolution to a dilemma.
Candidates conflate the regulatory domain with the ethics domain and assume the exam rewards minimum-compliance reasoning. Items frequently describe conduct that is permitted by law but violates reasonable expectations, dignity, or proportionality.
What to do instead
When an option rests only on legality, ask: 'Would this survive a stakeholder justification test?' Treat compliance as a floor, not a ceiling. In practice, write two columns - 'lawful?' and 'defensible to affected people?' - and reject options that pass the first and fail the second.
Diagnosing bias only at the model-output stage, ignoring proxy variables, labelling practices, sampling frames, and historical target construction.
Fairness teaching materials foreground metrics (demographic parity, equalised odds), so candidates jump to threshold adjustment. Exam items often place the defect upstream - in who was labelled, how the ground truth was defined, or which population was under-sampled.
What to do instead
Walk a fixed pipeline checklist for every bias scenario: problem formulation, population and sampling frame, label definition and annotator effects, feature set and proxies, model, evaluation subgroups, deployment context, feedback loops. Name the earliest stage where the harm originates, since the correct answer usually targets that stage.
Offering 'make the model explainable' as a complete answer to transparency questions without specifying to whom, at what level of abstraction, and for what decision purpose.
Explainability is taught as a single property, so candidates miss that a regulator, an affected individual, and an internal engineer need different artefacts (global feature attribution, local counterfactual, model documentation, audit access).
What to do instead
Always decompose a transparency item into audience, decision, and artefact. Match: affected individual contesting a decision → local, actionable reason; auditor → documentation and access; developer → global and diagnostic. Reject options that give a technically true explanation that the stated audience cannot use.
Assigning accountability to a role by seniority rather than by who held decision rights, control, and ability to have acted differently in the described workflow.
Candidates reason 'the executive is ultimately responsible', which is often a distractor. Accountability items test the distinction between causal contribution, role obligation, and capacity to prevent harm, including the 'many hands' problem in automated pipelines.
What to do instead
For accountability items, map the described workflow and ask at each node: who authorised, who could have stopped it, who bore the duty to check? Prefer answers that distribute obligations with named mechanisms (sign-off, review gates, audit rights) over answers that simply name a senior person.
Ignoring the territorial and jurisdictional dimension of data-protection scenarios, such as transfers, applicable law, or the difference between controller and processor obligations.
Candidates focus on the substantive privacy harm and skip the structural question of which regime and which role applies. Cross-border items are then answered by intuition rather than by role analysis.
What to do instead
Before answering, classify each party as controller, processor, or joint controller based on who determines purpose and means, then identify the transfer and the safeguard mechanism. Build a one-page comparison of the major regimes' transfer tools and breach-notification clocks so role and jurisdiction questions become mechanical.
Resolving whistleblowing scenarios by recommending immediate external disclosure, skipping internal channels, proportionality, and the documented escalation sequence.
Candidates admire whistleblowers and equate courage with going public fast. Exam items reward a graded response: raise internally, document, escalate, use the regulator, disclose publicly only when harm is serious, imminent, and internal channels have failed or are compromised.
What to do instead
Use a fixed ladder for every whistleblowing item: verify facts, use internal reporting and designated channels, document contemporaneously, escalate to regulator or prescribed body, and treat external or public disclosure as a last resort justified by severity, urgency, and futility of prior steps. Check whether the option protects the reporter from retaliation, since that is often the discriminator.
Describing emerging-technology impacts in generic hype or doom language instead of grounding the answer in a specific mechanism, affected population, and mitigations.
Generative models, synthetic data, and automated decision systems attract vague vocabulary, and candidates write what sounds sophisticated. Items ask for the concrete failure pathway - memorisation and leakage, model collapse from recursive training data, consent in scraped corpora, provenance loss.
What to do instead
For each emerging-technology topic, prepare a three-part note: mechanism (how the harm arises technically), affected party, and governance lever (provenance, licensing, evaluation, disclosure, human oversight). Answer using those three parts so the option you pick names a real mechanism rather than a mood.
Treating an ethical decision-making framework as a ritual to recite rather than a filter that eliminates options, so the chosen answer contradicts the framework just described.
Candidates list steps (identify stakeholders, weigh harms, decide) as a memorised paragraph and then answer from gut feeling, producing internally inconsistent responses. Case-analysis items are built to punish this gap.
What to do instead
Practise applying one framework end-to-end to full case studies under time pressure, writing the elimination rationale for each rejected option. Ensure your final answer is the one your stated framework would select, and check that the option addresses the most affected stakeholder rather than the most powerful one.