CDEP cheat sheet

Data ethics = systematic reflection on right/wrong in collection, processing, and use of data; distinct from law (legal minimum) and from data quality.

Four classical principles (bioethics-derived): beneficence (do good), non-maleficence (avoid harm), autonomy (respect agency/consent), justice (fair distribution of benefits/burdens).

Value-based framework: identify stakeholders, list affected values (privacy, fairness, autonomy, trust), and resolve conflicts explicitly before deployment.

Consequentialist test: weigh aggregate benefits vs harms; deontological test: check duties/rights (e.g., honesty, non-deception) regardless of outcome.

Virtue ethics test: would a prudent, honest, fair practitioner act this way? Useful when rules are silent.

Data ethics is context-dependent: same datum (e.g., location) may be low-risk in one use and high-risk in another; risk follows use, not just field.

Ethical decision procedure: (1) state the use; (2) identify affected parties; (3) map harms/benefits; (4) check legal floor; (5) check rights/consent; (6) document rationale.

Ethics vs. compliance: compliance asks 'may we?', ethics asks 'should we?'; lawful action can still be unethical (e.g., legal but manipulative profiling).

Transparency principle: affected persons should be able to know that data about them is used, for what purpose, and with what consequences.

Informed consent requires four elements: competence (capacity), disclosure (purposes, risks, uses), voluntariness (no coercion), and comprehension.

Purpose limitation: data collected for purpose A should not be repurposed for unrelated purpose B without fresh consent or a lawful, ethical justification.

Data minimization: collect only fields necessary for the stated purpose; each extra field adds breach and misuse risk without adding value.

Secondary use rule: re-use for research/analytics requires de-identification or an ethics review when re-consent is impractical.

Dark patterns (pre-checked boxes, confusing toggles, nagging) invalidate consent because they undermine voluntariness and comprehension.

Proxy variables (ZIP code, browsing history) can encode protected traits; ethical collection requires checking whether a proxy is a sensitive attribute in disguise.

Third-party sharing: disclose recipients and onward-transfer rules; silent resale of personal data breaches purpose limitation.

Withdrawal must be as easy as consent; data already used in models may be practically irremovable, so flag this at collection time.

Synthetic data is not automatically ethical: it can leak training records (membership inference) and can inherit the biases of its source.

Personal data = any information relating to an identified or identifiable natural person; identifiability is relative to available means (singling out, linkability, inference).

Anonymization vs. pseudonymization: pseudonymized data remains personal (key exists); anonymized data must be irreversible under reasonable re-identification effort.

GDPR core duties: lawful basis, purpose limitation, data minimization, accuracy, storage limitation, integrity/confidentiality, accountability.

Data subject rights (GDPR): access, rectification, erasure, restriction, portability, objection, and rights around automated decision-making.

Breach notification: GDPR requires notifying the supervisory authority within 72 hours of awareness where risk exists; notify data subjects when high risk.

CIA triad: confidentiality (access control, encryption), integrity (hashing, checksums, audit logs), availability (backups, redundancy).

Defense in depth: least privilege, MFA, encryption at rest and in transit, segmentation, logging, and regular patching are layered, not alternatives.

Cross-border transfer requires a lawful mechanism (adequacy decision, standard contractual clauses, binding corporate rules) plus a transfer impact assessment.

Privacy by design/default: embed protections in system architecture and default settings (e.g., opt-in sharing, minimal retention) rather than bolting them on.

Data governance roles: data owner (accountable), data steward (quality/custody), data custodian (technical controls), and a cross-functional ethics review board.

Model risk management: inventory models, assign risk tiers, validate independently, monitor drift, and require human sign-off for high-impact uses.

Bias taxonomy: historical bias (world reflects past inequity), sampling/selection bias, measurement bias, label bias, and deployment/feedback bias.

Fairness metrics conflict: demographic parity (equal positive rates), equalized odds (equal TPR/FPR across groups), and calibration cannot all hold simultaneously in general.

Proxy discrimination: a neutral feature (e.g., postal code) can reproduce protected-class disparity; test disparate impact, not just intent.

Explainability vs. accuracy trade-off: high-impact decisions (credit, hiring, health) generally require interpretable or post-hoc explained models.

Emerging-tech ethics: generative AI needs provenance/watermarking and hallucination controls; autonomous systems need meaningful human oversight; facial recognition needs necessity and proportionality tests.

Accountability must be assignable to a named human or role; 'the algorithm decided' is not an acceptable governance answer.

Audit trail: log data lineage, model versions, decisions, and overrides so outcomes can be reconstructed and challenged.