CDEI cheat sheet

Data ethics concerns the moral principles and values that govern how data is collected, processed, shared, and used, distinct from data law and data governance.

Four classical ethical principles for data (borrowed from bioethics): beneficence, non-maleficence, autonomy (respect for persons), and justice (fairness).

Moor's law of ethical lag: the rate of technological change outpaces the rate at which ethical norms and policies develop.

The Collingridge dilemma: early in a technology's development, its impacts are unknown; once known, the technology is entrenched and hard to change.

Data ethics extends beyond compliance: an action can be legal yet unethical (e.g., exploiting a legal loophole to infer sensitive attributes).

Key ethical concerns include privacy, autonomy, fairness, transparency, accountability, and the power asymmetries created by data concentration.

Ethical decision-making requires identifying stakeholders (data subjects, users, third parties, society) and assessing harms and benefits to each.

A common decision rule: when a data use would violate a reasonable person's expectation of privacy, treat it as presumptively unethical absent strong justification.

Data ethics is contextual: the same data practice may be acceptable in one setting (public health surveillance) and unacceptable in another (marketing).

Consequentialism (utilitarianism): an action is right if it produces the best overall consequences; evaluate by summing benefits and harms across stakeholders.

Deontology (Kantian ethics): some actions are inherently right or wrong regardless of outcomes; treat people as ends, never merely as means.

Virtue ethics: focus on the character traits (honesty, integrity, care) that a virtuous data practitioner would exhibit.

Justice as fairness (Rawls): design systems behind a 'veil of ignorance' so inequalities benefit the least advantaged (difference principle).

Ethics of care: emphasizes relationships, dependency, and the concrete responsibilities we have toward vulnerable parties.

Capability approach (Sen, Nussbaum): evaluate data practices by whether they expand or restrict people's real opportunities to live valued lives.

Value alignment: the goal of designing AI systems whose behavior reflects the intended values of their users and society.

When frameworks conflict, use a structured deliberation: identify the conflict, weigh the principles, document the rationale, and seek the least harmful resolution.

A useful heuristic: deontology sets side-constraints (never deceive, never exploit), while consequentialism guides trade-offs among permitted options.

Privacy is the individual's control over the collection, use, and disclosure of their personal information; confidentiality is the obligation to protect information shared in trust.

The Fair Information Practice Principles (FIPPs): notice, choice/consent, access, integrity/security, and enforcement.

k-anonymity: a record is indistinguishable from at least k-1 other records on the quasi-identifiers; it does not protect against attribute disclosure.

l-diversity: each equivalence class has at least l distinct values of the sensitive attribute, addressing homogeneity attacks on k-anonymity.

Differential privacy: a randomized mechanism satisfies ε-differential privacy if the output distributions on neighboring datasets differ by at most a factor of e^ε; smaller ε means stronger privacy.

Re-identification risk: combining de-identified data with auxiliary data can re-identify individuals (Netflix Prize and AOL search-log cases).

Data minimization: collect only the personal data necessary for a specified purpose; retain it only as long as needed.

Consent must be informed, specific, freely given, and revocable to be ethically valid; bundled or coerced consent is invalid.

Privacy by design: embed privacy protections into system architecture and default settings rather than adding them after deployment.

Historical bias: data reflects past inequities (e.g., biased policing records) and reproduces them when used for training.

Representation bias: the training sample underrepresents or misrepresents certain groups, leading to worse performance for them.

Measurement bias: a proxy variable is used for a construct and measures it differently across groups (e.g., arrest as a proxy for crime).

Aggregation bias: a single model is applied across heterogeneous subgroups when group-specific models would be more accurate.

Statistical parity (demographic parity): P(Ŷ=1|A=0) = P(Ŷ=1|A=1); equal positive prediction rates across groups.

Equalized odds: true positive and false positive rates are equal across groups; equal opportunity requires only equal true positive rates.

Predictive parity: equal positive predictive value across groups; calibration within groups.

The impossibility theorem: statistical parity, equalized odds, and predictive parity generally cannot all hold simultaneously when base rates differ.

Disparate impact: a facially neutral practice that disproportionately harms a protected group, regardless of intent (the 80% rule: selection rate ratio < 0.8).

Data governance: the system of decision rights, policies, and accountability mechanisms that govern data assets across their lifecycle.

Data stewardship: the operational responsibility for data quality, access, and appropriate use, often assigned to data stewards.

The data lifecycle: collection, processing, storage, sharing, retention, and disposal; controls should exist at each stage.

Data provenance: documentation of origin, lineage, and transformations; essential for auditability and trust.

A data catalog and metadata registry support discovery, ownership assignment, and access control.

Role separation: data owners (accountable for a domain), data stewards (operational), data custodians (technical), and data users.

Access control models: role-based (RBAC), attribute-based (ABAC), and purpose-based access; least-privilege principle applies.

Data retention schedules specify how long each data class is kept and how it is securely destroyed; legal holds override deletion.

A data protection impact assessment (DPIA) is required for high-risk processing under GDPR and is a governance control for identifying and mitigating risk.

Transparency: disclosing what data is used, how a system works, and what its limitations are; distinct from explainability.

Explainability: the ability to provide understandable reasons for a specific model output (local) or the model's overall behavior (global).

Intrinsic interpretability: models such as linear regression, decision trees, and generalized additive models are interpretable by design.

Post-hoc explanation methods: LIME, SHAP, counterfactual explanations, and feature importance approximate reasons after training.

The right to explanation under GDPR (Arts. 13-15, 22) concerns automated decision-making and meaningful information about the logic involved.

Accountability: the obligation to answer for decisions and their consequences; requires identifiable responsible parties and remedy mechanisms.

Algorithmic impact assessments and model cards document intended use, performance across groups, and known limitations.

Auditability requires logging of inputs, outputs, model versions, and decisions so that behavior can be reconstructed and reviewed.

A decision rule: the greater the potential harm of a decision, the stronger the required level of explainability and human oversight.

GDPR principles: lawfulness, fairness, and transparency; purpose limitation; data minimization; accuracy; storage limitation; integrity and confidentiality; accountability.

GDPR lawful bases: consent, contract, legal obligation, vital interests, public task, legitimate interests; special-category data needs an additional condition.

GDPR data subject rights: access, rectification, erasure, restriction, portability, objection, and rights related to automated decision-making.

CCPA/CPRA (California): rights to know, delete, correct, opt out of sale/sharing, and limit use of sensitive personal information; applies to businesses meeting thresholds.

HIPAA (US health data): Privacy Rule, Security Rule, breach notification; protected health information (PHI) and the minimum necessary standard.

FERPA (US education records), GLBA (financial data), and COPPA (children under 13) are sectoral US privacy laws.

The EU AI Act uses a risk-based approach: prohibited practices, high-risk systems (with conformity assessment), limited-risk (transparency), and minimal-risk.

The OECD AI Principles and UNESCO Recommendation on the Ethics of AI are non-binding international frameworks for trustworthy AI.

Cross-border data transfers require a lawful transfer mechanism (e.g., adequacy decision, standard contractual clauses, binding corporate rules).

Human-in-the-loop: a human reviews and can override consequential automated decisions; meaningful oversight requires authority, competence, and time.

Model risk management: validate models for accuracy, stability, fairness, and fitness for intended use before and after deployment.

Fairness testing should be performed on disaggregated subgroups, including intersectional groups, not just overall accuracy.

Training data provenance and consent: using data for a purpose beyond the original consent (function creep) is an ethical and often legal violation.

Synthetic data can reduce privacy risk but can inherit and amplify biases from the source data.

Monitoring for distribution shift and performance degradation after deployment is an ethical obligation, not only a technical one.

Red-teaming and adversarial testing identify harmful failure modes before release.

Deployment decision rule: if a model cannot be validated for fairness on relevant subgroups, restrict its use to low-stakes or research contexts.

Incident response: maintain a process for reporting, investigating, and remediating harms caused by AI systems.