CDEF cheat sheet

Consequentialism: an act is right if its outcomes maximize total welfare; in data work this maps to harm-benefit analysis and cost-benefit review of a model or data release.

Deontology: duties and rules bind regardless of outcome; consent, promise-keeping, and non-deception are constraints that a favorable aggregate result cannot override.

Virtue ethics: evaluates the agent's character (honesty, care, integrity) rather than the act or outcome; used to justify professional dispositions in data practice.

Justice as fairness (Rawls): principles chosen behind a veil of ignorance; supports evaluating a system by how it treats the least advantaged group.

Principlism (Beauchamp and Childress): four mid-level principles - beneficence, non-maleficence, autonomy, justice - commonly extended to data ethics as respect for autonomy, non-harm, beneficence, and fairness.

Moral relativism vs. universalism: relativism ties judgments to culture or context, universalism holds some norms apply across contexts; data ethics codes generally assume a universalist floor.

Fact-value distinction (Hume's is-ought gap): descriptive claims about data do not by themselves entail normative conclusions; a value premise must be supplied explicitly.

Distributive justice criteria: equality (same to all), equity (proportional to need or contribution), and priority to the worst-off yield different allocation rules for data benefits and burdens.

Ethical pluralism: multiple irreducible values (privacy, accuracy, autonomy, welfare) can conflict; resolution requires explicit trade-off justification, not a single metric.

Personal data: any information relating to an identified or identifiable natural person; identifiability can be direct (name) or indirect (a combination of quasi-identifiers).

Data protection principles (GDPR Art. 5): lawfulness, fairness, transparency; purpose limitation; data minimization; accuracy; storage limitation; integrity and confidentiality; accountability.

Legal bases for processing (GDPR Art. 6): consent, contract, legal obligation, vital interests, public task, legitimate interests; special-category data (Art. 9) requires an additional condition.

Data subject rights: access, rectification, erasure, restriction, portability, objection, and rights relating to automated decision-making; erasure is not absolute and yields to legal retention duties.

Consent validity: freely given, specific, informed, unambiguous; silence, pre-ticked boxes, and bundled terms do not constitute valid consent.

Anonymization vs. pseudonymization: pseudonymized data remain personal data because re-identification is possible with additional information held separately; anonymized data fall outside data protection scope.

k-anonymity: each record is indistinguishable from at least k-1 others on quasi-identifiers; it does not protect against attribute disclosure or homogeneity attacks.

Differential privacy: a randomized mechanism satisfies epsilon-differential privacy if outputs for neighboring datasets differ by at most a factor of e^epsilon; smaller epsilon means stronger privacy.

Privacy by design and by default: data protection is embedded in processing from the outset, and the most privacy-protective settings apply without user action.

Statistical parity (demographic parity): the positive decision rate is equal across groups; it ignores base rates and can conflict with predictive accuracy.

Equalized odds: true positive and false positive rates are equal across groups; equal opportunity is the weaker version requiring only equal true positive rates.

Predictive parity (calibration): among those receiving the same score, the outcome rate is equal across groups; calibration within groups and equalized odds generally cannot both hold when base rates differ.

Impossibility result (Chouldechova; Kleinberg et al.): with unequal base rates, calibration, balance for the positive class, and balance for the negative class cannot all be satisfied simultaneously.

Bias sources: historical bias in labels, representation bias in sampling, measurement bias in features, aggregation bias from treating heterogeneous groups as one, and deployment bias from use context.

Proxy variables: features correlated with a protected attribute (postcode, name, school) can transmit disparate impact even when the protected attribute is excluded.

Disparate impact (four-fifths or 80% rule): selection rate for a protected group below 80% of the highest group's rate is a screening signal of adverse impact.

Fairness metric selection rule: the metric must follow from the policy question - allocation decisions favor parity metrics, risk-scoring decisions favor calibration and equalized odds.

Mitigation stages: pre-processing (reweighting, resampling), in-processing (constrained optimization, adversarial debiasing), post-processing (group-specific thresholds).

Transparency types: disclosure of the fact of processing, of the data used, of model logic, and of performance; each answers a different question and satisfies different obligations.

GDPR Art. 13-15: controllers must provide meaningful information about the logic involved in automated decision-making, plus the significance and envisaged consequences.

GDPR Art. 22: data subjects have the right not to be subject to solely automated decisions producing legal or similarly significant effects, subject to exceptions (contract, law, explicit consent).

Interpretability vs. explainability: interpretability is an intrinsic property of simple models; explainability is post-hoc reconstruction of reasons for a complex model's output.

Local vs. global explanation: local explains a single prediction (LIME, SHAP values), global explains overall model behavior (feature importance, surrogate trees).

Explanation quality criteria: faithfulness to the model, stability under small input changes, comprehensibility for the intended audience, and completeness relative to the decision.

Accountability: the obligation to answer for an action; requires an identifiable actor, a documented standard, and a remedy path - transparency alone does not assign responsibility.

Algorithmic impact assessment: structured review of purpose, data, stakeholders, risks, mitigations, and monitoring before deployment, with periodic re-review.

Contestability: affected persons can challenge a decision and obtain a reasoned response; distinct from explanation, which only describes the basis of the decision.

Data governance roles: data owner (accountable for a domain), data steward (quality and policy execution), data custodian (technical safeguards), and the data protection officer (independent oversight).

Data lifecycle stages: acquisition, storage, processing, sharing, retention, archiving, and disposal; controls and ethical duties attach at every stage, not only collection.

CIA triad: confidentiality, integrity, availability; ethics adds accountability and non-maleficence as governance objectives alongside the triad.

Least privilege and need-to-know: access is limited to the minimum necessary for the task and reviewed periodically; role-based access control implements this at scale.

Data classification tiers (public, internal, confidential, restricted) drive handling rules; misclassification is a governance failure even absent a breach.

Retention and deletion: storage limitation requires defined retention periods and verifiable deletion; backups and derived datasets must be included in deletion scope.

Breach notification (GDPR Art. 33-34): notify the supervisory authority within 72 hours of awareness unless unlikely to result in risk; notify subjects when there is high risk.

Security ethics: proportionality of monitoring, no secondary use of security telemetry without a lawful basis, and disclosure of vulnerabilities handled through coordinated disclosure.

GDPR territorial scope (Art. 3): applies to processing in the EU by an establishment, and to non-EU controllers offering goods or services to, or monitoring, EU data subjects.

GDPR penalties: up to EUR 10 million or 2% of global annual turnover for certain obligations; up to EUR 20 million or 4% for core principles and rights, whichever is higher.

Roles: controller determines purposes and means; processor acts on the controller's instructions; a joint-controller arrangement requires an agreed allocation of responsibilities.

DPIA (Art. 35): required for high-risk processing, including large-scale systematic monitoring and large-scale special-category data; consult the supervisory authority if residual risk remains high.

EU AI Act risk tiers: prohibited practices, high-risk systems (conformity assessment, registration, human oversight), limited-risk (transparency duties), minimal-risk (no added obligations).

Sectoral regimes: HIPAA governs protected health information in the US; GLBA covers financial data; COPPA restricts collection from children under 13 in the US.

Cross-border transfers (GDPR Chapter V): adequacy decision, standard contractual clauses, binding corporate rules, or derogations; Schrems II requires a transfer impact assessment.

Legal vs. ethical compliance: lawful processing can still be unethical; regulation sets a floor, and ethical review addresses harms regulation does not reach.

Stakeholder mapping: identify who is affected, the nature and magnitude of impact, power and dependency, and whether affected parties are represented in the decision.

Structured test sequence: facts, stakeholders, options, consequences, duties and rights, virtues, and a final justification that names the trade-off accepted.

Precautionary principle: where harm is plausible and uncertain, take proportionate preventive action rather than waiting for conclusive evidence.

Reversibility test: prefer options whose harms can be undone; irreversible actions (public release, model deployment at scale) demand a higher evidentiary bar.

Harm matrix: rate each option on severity, likelihood, reversibility, and number affected; the product guides but does not replace judgment.

Sunlight (publicity) test: would the decision withstand disclosure to affected users and to peers? Inability to defend it publicly is a warning signal.

Moral residue: after a justified trade-off, some obligation remains unmet; record it and mitigate it rather than declaring the conflict fully resolved.

Case analysis discipline: separate the descriptive question (what does the system do) from the normative question (what should be done), and state the value premise explicitly.

Escalation rule: when a decision exceeds the analyst's authority or the risk is severe and irreversible, refer upward before acting.

Generative models: trained on large corpora and can emit memorized or near-verbatim training content; memorization risk rises with duplication and model capacity.

Training-data provenance: licensing, consent, and copyright status of training corpora are unresolved in many jurisdictions and must be documented for defensibility.

Synthetic data: can reduce exposure of real personal data, but inherits and can amplify biases of the source distribution and may be re-identifiable through membership inference.

Federated learning: model updates rather than raw data leave the device; gradients can leak training examples, so secure aggregation and differential privacy are layered on.

Model drift: performance degrades as the deployment distribution shifts; monitoring for drift is an ethical obligation, not only an engineering one.

Automated decision systems in high-stakes domains (credit, hiring, health, justice) attract stricter scrutiny because errors are consequential and hard to reverse.

Environmental cost: training and inference consume energy and water; sustainability is an emerging ethical consideration in model selection.

Deepfakes and synthetic media: provenance standards (content credentials, watermarking) and disclosure duties are the main mitigation levers.

Dual-use risk: the same capability can serve beneficial and harmful ends; release decisions should weigh uplift to malicious actors against scientific benefit.

Psychological safety: team members can raise concerns without fear of punishment; it is a precondition for surfacing ethical risk early.

Professional codes: ACM Code of Ethics and IEEE Ethically Aligned Design emphasize avoiding harm, honesty, fairness, and respecting privacy and confidentiality.

Conflict of interest: a personal interest that could improperly influence professional judgment; the standard remedy is disclosure plus recusal from the decision.

Whistleblowing channels: internal reporting first, then regulator or designated external body; many jurisdictions protect disclosures made through prescribed channels.

Protected disclosure conditions: the concern is a genuine legal or ethical breach, the disclosure is proportionate, and internal routes were exhausted or demonstrably unsafe.

Retaliation prohibition: adverse action against a good-faith reporter is itself a violation; good faith protects the reporter even if the concern proves unfounded.

Documentation duty: contemporaneous records of concerns raised, responses received, and decisions made are the evidentiary basis for any later claim.

Ethics review boards: independent bodies that review high-risk data projects; their authority depends on a mandate to halt work, not merely to advise.

Tone at the top: leadership behavior, not published policy, is the strongest predictor of whether ethical concerns are reported internally.

Consequentialism: an act is right if its outcomes maximize total welfare; in data work this maps to harm-benefit analysis and cost-benefit review of a model or data release.

Deontology: duties and rules bind regardless of outcome; consent, promise-keeping, and non-deception are constraints that a favorable aggregate result cannot override.

Virtue ethics: evaluates the agent's character (honesty, care, integrity) rather than the act or outcome; used to justify professional dispositions in data practice.

Justice as fairness (Rawls): principles chosen behind a veil of ignorance; supports evaluating a system by how it treats the least advantaged group.

Principlism (Beauchamp and Childress): four mid-level principles - beneficence, non-maleficence, autonomy, justice - commonly extended to data ethics as respect for autonomy, non-harm, beneficence, and fairness.

Moral relativism vs. universalism: relativism ties judgments to culture or context, universalism holds some norms apply across contexts; data ethics codes generally assume a universalist floor.

Fact-value distinction (Hume's is-ought gap): descriptive claims about data do not by themselves entail normative conclusions; a value premise must be supplied explicitly.

Distributive justice criteria: equality (same to all), equity (proportional to need or contribution), and priority to the worst-off yield different allocation rules for data benefits and burdens.

Ethical pluralism: multiple irreducible values (privacy, accuracy, autonomy, welfare) can conflict; resolution requires explicit trade-off justification, not a single metric.

Personal data: any information relating to an identified or identifiable natural person; identifiability can be direct (name) or indirect (a combination of quasi-identifiers).

Data protection principles (GDPR Art. 5): lawfulness, fairness, transparency; purpose limitation; data minimization; accuracy; storage limitation; integrity and confidentiality; accountability.

Legal bases for processing (GDPR Art. 6): consent, contract, legal obligation, vital interests, public task, legitimate interests; special-category data (Art. 9) requires an additional condition.

Data subject rights: access, rectification, erasure, restriction, portability, objection, and rights relating to automated decision-making; erasure is not absolute and yields to legal retention duties.

Consent validity: freely given, specific, informed, unambiguous; silence, pre-ticked boxes, and bundled terms do not constitute valid consent.

Anonymization vs. pseudonymization: pseudonymized data remain personal data because re-identification is possible with additional information held separately; anonymized data fall outside data protection scope.

k-anonymity: each record is indistinguishable from at least k-1 others on quasi-identifiers; it does not protect against attribute disclosure or homogeneity attacks.

Differential privacy: a randomized mechanism satisfies epsilon-differential privacy if outputs for neighboring datasets differ by at most a factor of e^epsilon; smaller epsilon means stronger privacy.

Privacy by design and by default: data protection is embedded in processing from the outset, and the most privacy-protective settings apply without user action.

Statistical parity (demographic parity): the positive decision rate is equal across groups; it ignores base rates and can conflict with predictive accuracy.

Equalized odds: true positive and false positive rates are equal across groups; equal opportunity is the weaker version requiring only equal true positive rates.

Predictive parity (calibration): among those receiving the same score, the outcome rate is equal across groups; calibration within groups and equalized odds generally cannot both hold when base rates differ.

Impossibility result (Chouldechova; Kleinberg et al.): with unequal base rates, calibration, balance for the positive class, and balance for the negative class cannot all be satisfied simultaneously.

Bias sources: historical bias in labels, representation bias in sampling, measurement bias in features, aggregation bias from treating heterogeneous groups as one, and deployment bias from use context.

Proxy variables: features correlated with a protected attribute (postcode, name, school) can transmit disparate impact even when the protected attribute is excluded.

Disparate impact (four-fifths or 80% rule): selection rate for a protected group below 80% of the highest group's rate is a screening signal of adverse impact.

Fairness metric selection rule: the metric must follow from the policy question - allocation decisions favor parity metrics, risk-scoring decisions favor calibration and equalized odds.

Mitigation stages: pre-processing (reweighting, resampling), in-processing (constrained optimization, adversarial debiasing), post-processing (group-specific thresholds).

Transparency types: disclosure of the fact of processing, of the data used, of model logic, and of performance; each answers a different question and satisfies different obligations.

GDPR Art. 13-15: controllers must provide meaningful information about the logic involved in automated decision-making, plus the significance and envisaged consequences.

GDPR Art. 22: data subjects have the right not to be subject to solely automated decisions producing legal or similarly significant effects, subject to exceptions (contract, law, explicit consent).

Interpretability vs. explainability: interpretability is an intrinsic property of simple models; explainability is post-hoc reconstruction of reasons for a complex model's output.

Local vs. global explanation: local explains a single prediction (LIME, SHAP values), global explains overall model behavior (feature importance, surrogate trees).

Explanation quality criteria: faithfulness to the model, stability under small input changes, comprehensibility for the intended audience, and completeness relative to the decision.

Accountability: the obligation to answer for an action; requires an identifiable actor, a documented standard, and a remedy path - transparency alone does not assign responsibility.

Algorithmic impact assessment: structured review of purpose, data, stakeholders, risks, mitigations, and monitoring before deployment, with periodic re-review.

Contestability: affected persons can challenge a decision and obtain a reasoned response; distinct from explanation, which only describes the basis of the decision.

Data governance roles: data owner (accountable for a domain), data steward (quality and policy execution), data custodian (technical safeguards), and the data protection officer (independent oversight).

Data lifecycle stages: acquisition, storage, processing, sharing, retention, archiving, and disposal; controls and ethical duties attach at every stage, not only collection.

CIA triad: confidentiality, integrity, availability; ethics adds accountability and non-maleficence as governance objectives alongside the triad.

Least privilege and need-to-know: access is limited to the minimum necessary for the task and reviewed periodically; role-based access control implements this at scale.

Data classification tiers (public, internal, confidential, restricted) drive handling rules; misclassification is a governance failure even absent a breach.

Retention and deletion: storage limitation requires defined retention periods and verifiable deletion; backups and derived datasets must be included in deletion scope.

Breach notification (GDPR Art. 33-34): notify the supervisory authority within 72 hours of awareness unless unlikely to result in risk; notify subjects when there is high risk.

Security ethics: proportionality of monitoring, no secondary use of security telemetry without a lawful basis, and disclosure of vulnerabilities handled through coordinated disclosure.

GDPR territorial scope (Art. 3): applies to processing in the EU by an establishment, and to non-EU controllers offering goods or services to, or monitoring, EU data subjects.

GDPR penalties: up to EUR 10 million or 2% of global annual turnover for certain obligations; up to EUR 20 million or 4% for core principles and rights, whichever is higher.

Roles: controller determines purposes and means; processor acts on the controller's instructions; a joint-controller arrangement requires an agreed allocation of responsibilities.

DPIA (Art. 35): required for high-risk processing, including large-scale systematic monitoring and large-scale special-category data; consult the supervisory authority if residual risk remains high.

EU AI Act risk tiers: prohibited practices, high-risk systems (conformity assessment, registration, human oversight), limited-risk (transparency duties), minimal-risk (no added obligations).

Sectoral regimes: HIPAA governs protected health information in the US; GLBA covers financial data; COPPA restricts collection from children under 13 in the US.

Cross-border transfers (GDPR Chapter V): adequacy decision, standard contractual clauses, binding corporate rules, or derogations; Schrems II requires a transfer impact assessment.

Legal vs. ethical compliance: lawful processing can still be unethical; regulation sets a floor, and ethical review addresses harms regulation does not reach.

Stakeholder mapping: identify who is affected, the nature and magnitude of impact, power and dependency, and whether affected parties are represented in the decision.

Structured test sequence: facts, stakeholders, options, consequences, duties and rights, virtues, and a final justification that names the trade-off accepted.

Precautionary principle: where harm is plausible and uncertain, take proportionate preventive action rather than waiting for conclusive evidence.

Reversibility test: prefer options whose harms can be undone; irreversible actions (public release, model deployment at scale) demand a higher evidentiary bar.

Harm matrix: rate each option on severity, likelihood, reversibility, and number affected; the product guides but does not replace judgment.

Sunlight (publicity) test: would the decision withstand disclosure to affected users and to peers? Inability to defend it publicly is a warning signal.

Moral residue: after a justified trade-off, some obligation remains unmet; record it and mitigate it rather than declaring the conflict fully resolved.

Case analysis discipline: separate the descriptive question (what does the system do) from the normative question (what should be done), and state the value premise explicitly.

Escalation rule: when a decision exceeds the analyst's authority or the risk is severe and irreversible, refer upward before acting.

Generative models: trained on large corpora and can emit memorized or near-verbatim training content; memorization risk rises with duplication and model capacity.

Training-data provenance: licensing, consent, and copyright status of training corpora are unresolved in many jurisdictions and must be documented for defensibility.

Synthetic data: can reduce exposure of real personal data, but inherits and can amplify biases of the source distribution and may be re-identifiable through membership inference.

Federated learning: model updates rather than raw data leave the device; gradients can leak training examples, so secure aggregation and differential privacy are layered on.

Model drift: performance degrades as the deployment distribution shifts; monitoring for drift is an ethical obligation, not only an engineering one.

Automated decision systems in high-stakes domains (credit, hiring, health, justice) attract stricter scrutiny because errors are consequential and hard to reverse.

Environmental cost: training and inference consume energy and water; sustainability is an emerging ethical consideration in model selection.

Deepfakes and synthetic media: provenance standards (content credentials, watermarking) and disclosure duties are the main mitigation levers.

Dual-use risk: the same capability can serve beneficial and harmful ends; release decisions should weigh uplift to malicious actors against scientific benefit.

Psychological safety: team members can raise concerns without fear of punishment; it is a precondition for surfacing ethical risk early.

Professional codes: ACM Code of Ethics and IEEE Ethically Aligned Design emphasize avoiding harm, honesty, fairness, and respecting privacy and confidentiality.

Conflict of interest: a personal interest that could improperly influence professional judgment; the standard remedy is disclosure plus recusal from the decision.

Whistleblowing channels: internal reporting first, then regulator or designated external body; many jurisdictions protect disclosures made through prescribed channels.

Protected disclosure conditions: the concern is a genuine legal or ethical breach, the disclosure is proportionate, and internal routes were exhausted or demonstrably unsafe.

Retaliation prohibition: adverse action against a good-faith reporter is itself a violation; good faith protects the reporter even if the concern proves unfounded.

Documentation duty: contemporaneous records of concerns raised, responses received, and decisions made are the evidentiary basis for any later claim.

Ethics review boards: independent bodies that review high-risk data projects; their authority depends on a mandate to halt work, not merely to advise.

Tone at the top: leadership behavior, not published policy, is the strongest predictor of whether ethical concerns are reported internally.

Consequentialism: an act is right if its outcomes maximize total welfare; in data work this maps to harm-benefit analysis and cost-benefit review of a model or data release.

Deontology: duties and rules bind regardless of outcome; consent, promise-keeping, and non-deception are constraints that a favorable aggregate result cannot override.

Virtue ethics: evaluates the agent's character (honesty, care, integrity) rather than the act or outcome; used to justify professional dispositions in data practice.

Justice as fairness (Rawls): principles chosen behind a veil of ignorance; supports evaluating a system by how it treats the least advantaged group.

Principlism (Beauchamp and Childress): four mid-level principles - beneficence, non-maleficence, autonomy, justice - commonly extended to data ethics as respect for autonomy, non-harm, beneficence, and fairness.

Moral relativism vs. universalism: relativism ties judgments to culture or context, universalism holds some norms apply across contexts; data ethics codes generally assume a universalist floor.

Fact-value distinction (Hume's is-ought gap): descriptive claims about data do not by themselves entail normative conclusions; a value premise must be supplied explicitly.

Distributive justice criteria: equality (same to all), equity (proportional to need or contribution), and priority to the worst-off yield different allocation rules for data benefits and burdens.

Ethical pluralism: multiple irreducible values (privacy, accuracy, autonomy, welfare) can conflict; resolution requires explicit trade-off justification, not a single metric.

Personal data: any information relating to an identified or identifiable natural person; identifiability can be direct (name) or indirect (a combination of quasi-identifiers).

Data protection principles (GDPR Art. 5): lawfulness, fairness, transparency; purpose limitation; data minimization; accuracy; storage limitation; integrity and confidentiality; accountability.

Legal bases for processing (GDPR Art. 6): consent, contract, legal obligation, vital interests, public task, legitimate interests; special-category data (Art. 9) requires an additional condition.

Data subject rights: access, rectification, erasure, restriction, portability, objection, and rights relating to automated decision-making; erasure is not absolute and yields to legal retention duties.

Consent validity: freely given, specific, informed, unambiguous; silence, pre-ticked boxes, and bundled terms do not constitute valid consent.

Anonymization vs. pseudonymization: pseudonymized data remain personal data because re-identification is possible with additional information held separately; anonymized data fall outside data protection scope.

k-anonymity: each record is indistinguishable from at least k-1 others on quasi-identifiers; it does not protect against attribute disclosure or homogeneity attacks.

Differential privacy: a randomized mechanism satisfies epsilon-differential privacy if outputs for neighboring datasets differ by at most a factor of e^epsilon; smaller epsilon means stronger privacy.

Privacy by design and by default: data protection is embedded in processing from the outset, and the most privacy-protective settings apply without user action.

Statistical parity (demographic parity): the positive decision rate is equal across groups; it ignores base rates and can conflict with predictive accuracy.

Equalized odds: true positive and false positive rates are equal across groups; equal opportunity is the weaker version requiring only equal true positive rates.

Predictive parity (calibration): among those receiving the same score, the outcome rate is equal across groups; calibration within groups and equalized odds generally cannot both hold when base rates differ.

Impossibility result (Chouldechova; Kleinberg et al.): with unequal base rates, calibration, balance for the positive class, and balance for the negative class cannot all be satisfied simultaneously.

Bias sources: historical bias in labels, representation bias in sampling, measurement bias in features, aggregation bias from treating heterogeneous groups as one, and deployment bias from use context.

Proxy variables: features correlated with a protected attribute (postcode, name, school) can transmit disparate impact even when the protected attribute is excluded.

Disparate impact (four-fifths or 80% rule): selection rate for a protected group below 80% of the highest group's rate is a screening signal of adverse impact.

Fairness metric selection rule: the metric must follow from the policy question - allocation decisions favor parity metrics, risk-scoring decisions favor calibration and equalized odds.

Mitigation stages: pre-processing (reweighting, resampling), in-processing (constrained optimization, adversarial debiasing), post-processing (group-specific thresholds).

Transparency types: disclosure of the fact of processing, of the data used, of model logic, and of performance; each answers a different question and satisfies different obligations.

GDPR Art. 13-15: controllers must provide meaningful information about the logic involved in automated decision-making, plus the significance and envisaged consequences.

GDPR Art. 22: data subjects have the right not to be subject to solely automated decisions producing legal or similarly significant effects, subject to exceptions (contract, law, explicit consent).

Interpretability vs. explainability: interpretability is an intrinsic property of simple models; explainability is post-hoc reconstruction of reasons for a complex model's output.

Local vs. global explanation: local explains a single prediction (LIME, SHAP values), global explains overall model behavior (feature importance, surrogate trees).

Explanation quality criteria: faithfulness to the model, stability under small input changes, comprehensibility for the intended audience, and completeness relative to the decision.

Accountability: the obligation to answer for an action; requires an identifiable actor, a documented standard, and a remedy path - transparency alone does not assign responsibility.

Algorithmic impact assessment: structured review of purpose, data, stakeholders, risks, mitigations, and monitoring before deployment, with periodic re-review.

Contestability: affected persons can challenge a decision and obtain a reasoned response; distinct from explanation, which only describes the basis of the decision.

Data governance roles: data owner (accountable for a domain), data steward (quality and policy execution), data custodian (technical safeguards), and the data protection officer (independent oversight).

Data lifecycle stages: acquisition, storage, processing, sharing, retention, archiving, and disposal; controls and ethical duties attach at every stage, not only collection.

CIA triad: confidentiality, integrity, availability; ethics adds accountability and non-maleficence as governance objectives alongside the triad.

Least privilege and need-to-know: access is limited to the minimum necessary for the task and reviewed periodically; role-based access control implements this at scale.

Data classification tiers (public, internal, confidential, restricted) drive handling rules; misclassification is a governance failure even absent a breach.

Retention and deletion: storage limitation requires defined retention periods and verifiable deletion; backups and derived datasets must be included in deletion scope.

Breach notification (GDPR Art. 33-34): notify the supervisory authority within 72 hours of awareness unless unlikely to result in risk; notify subjects when there is high risk.

Security ethics: proportionality of monitoring, no secondary use of security telemetry without a lawful basis, and disclosure of vulnerabilities handled through coordinated disclosure.

GDPR territorial scope (Art. 3): applies to processing in the EU by an establishment, and to non-EU controllers offering goods or services to, or monitoring, EU data subjects.

GDPR penalties: up to EUR 10 million or 2% of global annual turnover for certain obligations; up to EUR 20 million or 4% for core principles and rights, whichever is higher.

Roles: controller determines purposes and means; processor acts on the controller's instructions; a joint-controller arrangement requires an agreed allocation of responsibilities.

DPIA (Art. 35): required for high-risk processing, including large-scale systematic monitoring and large-scale special-category data; consult the supervisory authority if residual risk remains high.

EU AI Act risk tiers: prohibited practices, high-risk systems (conformity assessment, registration, human oversight), limited-risk (transparency duties), minimal-risk (no added obligations).

Sectoral regimes: HIPAA governs protected health information in the US; GLBA covers financial data; COPPA restricts collection from children under 13 in the US.

Cross-border transfers (GDPR Chapter V): adequacy decision, standard contractual clauses, binding corporate rules, or derogations; Schrems II requires a transfer impact assessment.

Legal vs. ethical compliance: lawful processing can still be unethical; regulation sets a floor, and ethical review addresses harms regulation does not reach.

Stakeholder mapping: identify who is affected, the nature and magnitude of impact, power and dependency, and whether affected parties are represented in the decision.

Structured test sequence: facts, stakeholders, options, consequences, duties and rights, virtues, and a final justification that names the trade-off accepted.

Precautionary principle: where harm is plausible and uncertain, take proportionate preventive action rather than waiting for conclusive evidence.

Reversibility test: prefer options whose harms can be undone; irreversible actions (public release, model deployment at scale) demand a higher evidentiary bar.

Harm matrix: rate each option on severity, likelihood, reversibility, and number affected; the product guides but does not replace judgment.

Sunlight (publicity) test: would the decision withstand disclosure to affected users and to peers? Inability to defend it publicly is a warning signal.

Moral residue: after a justified trade-off, some obligation remains unmet; record it and mitigate it rather than declaring the conflict fully resolved.

Case analysis discipline: separate the descriptive question (what does the system do) from the normative question (what should be done), and state the value premise explicitly.

Escalation rule: when a decision exceeds the analyst's authority or the risk is severe and irreversible, refer upward before acting.

Generative models: trained on large corpora and can emit memorized or near-verbatim training content; memorization risk rises with duplication and model capacity.

Training-data provenance: licensing, consent, and copyright status of training corpora are unresolved in many jurisdictions and must be documented for defensibility.

Synthetic data: can reduce exposure of real personal data, but inherits and can amplify biases of the source distribution and may be re-identifiable through membership inference.

Federated learning: model updates rather than raw data leave the device; gradients can leak training examples, so secure aggregation and differential privacy are layered on.

Model drift: performance degrades as the deployment distribution shifts; monitoring for drift is an ethical obligation, not only an engineering one.

Automated decision systems in high-stakes domains (credit, hiring, health, justice) attract stricter scrutiny because errors are consequential and hard to reverse.

Environmental cost: training and inference consume energy and water; sustainability is an emerging ethical consideration in model selection.

Deepfakes and synthetic media: provenance standards (content credentials, watermarking) and disclosure duties are the main mitigation levers.

Dual-use risk: the same capability can serve beneficial and harmful ends; release decisions should weigh uplift to malicious actors against scientific benefit.

Psychological safety: team members can raise concerns without fear of punishment; it is a precondition for surfacing ethical risk early.

Professional codes: ACM Code of Ethics and IEEE Ethically Aligned Design emphasize avoiding harm, honesty, fairness, and respecting privacy and confidentiality.

Conflict of interest: a personal interest that could improperly influence professional judgment; the standard remedy is disclosure plus recusal from the decision.

Whistleblowing channels: internal reporting first, then regulator or designated external body; many jurisdictions protect disclosures made through prescribed channels.

Protected disclosure conditions: the concern is a genuine legal or ethical breach, the disclosure is proportionate, and internal routes were exhausted or demonstrably unsafe.

Retaliation prohibition: adverse action against a good-faith reporter is itself a violation; good faith protects the reporter even if the concern proves unfounded.

Documentation duty: contemporaneous records of concerns raised, responses received, and decisions made are the evidentiary basis for any later claim.

Ethics review boards: independent bodies that review high-risk data projects; their authority depends on a mandate to halt work, not merely to advise.

Tone at the top: leadership behavior, not published policy, is the strongest predictor of whether ethical concerns are reported internally.

Consequentialism: an act is right if its outcomes maximize total welfare; in data work this maps to harm-benefit analysis and cost-benefit review of a model or data release.

Deontology: duties and rules bind regardless of outcome; consent, promise-keeping, and non-deception are constraints that a favorable aggregate result cannot override.

Virtue ethics: evaluates the agent's character (honesty, care, integrity) rather than the act or outcome; used to justify professional dispositions in data practice.

Justice as fairness (Rawls): principles chosen behind a veil of ignorance; supports evaluating a system by how it treats the least advantaged group.

Principlism (Beauchamp and Childress): four mid-level principles - beneficence, non-maleficence, autonomy, justice - commonly extended to data ethics as respect for autonomy, non-harm, beneficence, and fairness.

Moral relativism vs. universalism: relativism ties judgments to culture or context, universalism holds some norms apply across contexts; data ethics codes generally assume a universalist floor.

Fact-value distinction (Hume's is-ought gap): descriptive claims about data do not by themselves entail normative conclusions; a value premise must be supplied explicitly.

Distributive justice criteria: equality (same to all), equity (proportional to need or contribution), and priority to the worst-off yield different allocation rules for data benefits and burdens.

Ethical pluralism: multiple irreducible values (privacy, accuracy, autonomy, welfare) can conflict; resolution requires explicit trade-off justification, not a single metric.

Personal data: any information relating to an identified or identifiable natural person; identifiability can be direct (name) or indirect (a combination of quasi-identifiers).

Data protection principles (GDPR Art. 5): lawfulness, fairness, transparency; purpose limitation; data minimization; accuracy; storage limitation; integrity and confidentiality; accountability.

Legal bases for processing (GDPR Art. 6): consent, contract, legal obligation, vital interests, public task, legitimate interests; special-category data (Art. 9) requires an additional condition.

Data subject rights: access, rectification, erasure, restriction, portability, objection, and rights relating to automated decision-making; erasure is not absolute and yields to legal retention duties.

Consent validity: freely given, specific, informed, unambiguous; silence, pre-ticked boxes, and bundled terms do not constitute valid consent.

Anonymization vs. pseudonymization: pseudonymized data remain personal data because re-identification is possible with additional information held separately; anonymized data fall outside data protection scope.

k-anonymity: each record is indistinguishable from at least k-1 others on quasi-identifiers; it does not protect against attribute disclosure or homogeneity attacks.

Differential privacy: a randomized mechanism satisfies epsilon-differential privacy if outputs for neighboring datasets differ by at most a factor of e^epsilon; smaller epsilon means stronger privacy.

Privacy by design and by default: data protection is embedded in processing from the outset, and the most privacy-protective settings apply without user action.

Statistical parity (demographic parity): the positive decision rate is equal across groups; it ignores base rates and can conflict with predictive accuracy.

Equalized odds: true positive and false positive rates are equal across groups; equal opportunity is the weaker version requiring only equal true positive rates.

Predictive parity (calibration): among those receiving the same score, the outcome rate is equal across groups; calibration within groups and equalized odds generally cannot both hold when base rates differ.

Impossibility result (Chouldechova; Kleinberg et al.): with unequal base rates, calibration, balance for the positive class, and balance for the negative class cannot all be satisfied simultaneously.

Bias sources: historical bias in labels, representation bias in sampling, measurement bias in features, aggregation bias from treating heterogeneous groups as one, and deployment bias from use context.

Proxy variables: features correlated with a protected attribute (postcode, name, school) can transmit disparate impact even when the protected attribute is excluded.

Disparate impact (four-fifths or 80% rule): selection rate for a protected group below 80% of the highest group's rate is a screening signal of adverse impact.

Fairness metric selection rule: the metric must follow from the policy question - allocation decisions favor parity metrics, risk-scoring decisions favor calibration and equalized odds.

Mitigation stages: pre-processing (reweighting, resampling), in-processing (constrained optimization, adversarial debiasing), post-processing (group-specific thresholds).

Transparency types: disclosure of the fact of processing, of the data used, of model logic, and of performance; each answers a different question and satisfies different obligations.

GDPR Art. 13-15: controllers must provide meaningful information about the logic involved in automated decision-making, plus the significance and envisaged consequences.

GDPR Art. 22: data subjects have the right not to be subject to solely automated decisions producing legal or similarly significant effects, subject to exceptions (contract, law, explicit consent).

Interpretability vs. explainability: interpretability is an intrinsic property of simple models; explainability is post-hoc reconstruction of reasons for a complex model's output.

Local vs. global explanation: local explains a single prediction (LIME, SHAP values), global explains overall model behavior (feature importance, surrogate trees).

Explanation quality criteria: faithfulness to the model, stability under small input changes, comprehensibility for the intended audience, and completeness relative to the decision.

Accountability: the obligation to answer for an action; requires an identifiable actor, a documented standard, and a remedy path - transparency alone does not assign responsibility.

Algorithmic impact assessment: structured review of purpose, data, stakeholders, risks, mitigations, and monitoring before deployment, with periodic re-review.

Contestability: affected persons can challenge a decision and obtain a reasoned response; distinct from explanation, which only describes the basis of the decision.

Data governance roles: data owner (accountable for a domain), data steward (quality and policy execution), data custodian (technical safeguards), and the data protection officer (independent oversight).

Data lifecycle stages: acquisition, storage, processing, sharing, retention, archiving, and disposal; controls and ethical duties attach at every stage, not only collection.

CIA triad: confidentiality, integrity, availability; ethics adds accountability and non-maleficence as governance objectives alongside the triad.

Least privilege and need-to-know: access is limited to the minimum necessary for the task and reviewed periodically; role-based access control implements this at scale.

Data classification tiers (public, internal, confidential, restricted) drive handling rules; misclassification is a governance failure even absent a breach.

Retention and deletion: storage limitation requires defined retention periods and verifiable deletion; backups and derived datasets must be included in deletion scope.

Breach notification (GDPR Art. 33-34): notify the supervisory authority within 72 hours of awareness unless unlikely to result in risk; notify subjects when there is high risk.

Security ethics: proportionality of monitoring, no secondary use of security telemetry without a lawful basis, and disclosure of vulnerabilities handled through coordinated disclosure.

GDPR territorial scope (Art. 3): applies to processing in the EU by an establishment, and to non-EU controllers offering goods or services to, or monitoring, EU data subjects.

GDPR penalties: up to EUR 10 million or 2% of global annual turnover for certain obligations; up to EUR 20 million or 4% for core principles and rights, whichever is higher.

Roles: controller determines purposes and means; processor acts on the controller's instructions; a joint-controller arrangement requires an agreed allocation of responsibilities.

DPIA (Art. 35): required for high-risk processing, including large-scale systematic monitoring and large-scale special-category data; consult the supervisory authority if residual risk remains high.

EU AI Act risk tiers: prohibited practices, high-risk systems (conformity assessment, registration, human oversight), limited-risk (transparency duties), minimal-risk (no added obligations).

Sectoral regimes: HIPAA governs protected health information in the US; GLBA covers financial data; COPPA restricts collection from children under 13 in the US.

Cross-border transfers (GDPR Chapter V): adequacy decision, standard contractual clauses, binding corporate rules, or derogations; Schrems II requires a transfer impact assessment.

Legal vs. ethical compliance: lawful processing can still be unethical; regulation sets a floor, and ethical review addresses harms regulation does not reach.

Stakeholder mapping: identify who is affected, the nature and magnitude of impact, power and dependency, and whether affected parties are represented in the decision.

Structured test sequence: facts, stakeholders, options, consequences, duties and rights, virtues, and a final justification that names the trade-off accepted.

Precautionary principle: where harm is plausible and uncertain, take proportionate preventive action rather than waiting for conclusive evidence.

Reversibility test: prefer options whose harms can be undone; irreversible actions (public release, model deployment at scale) demand a higher evidentiary bar.

Harm matrix: rate each option on severity, likelihood, reversibility, and number affected; the product guides but does not replace judgment.

Sunlight (publicity) test: would the decision withstand disclosure to affected users and to peers? Inability to defend it publicly is a warning signal.

Moral residue: after a justified trade-off, some obligation remains unmet; record it and mitigate it rather than declaring the conflict fully resolved.

Case analysis discipline: separate the descriptive question (what does the system do) from the normative question (what should be done), and state the value premise explicitly.

Escalation rule: when a decision exceeds the analyst's authority or the risk is severe and irreversible, refer upward before acting.

Generative models: trained on large corpora and can emit memorized or near-verbatim training content; memorization risk rises with duplication and model capacity.

Training-data provenance: licensing, consent, and copyright status of training corpora are unresolved in many jurisdictions and must be documented for defensibility.

Synthetic data: can reduce exposure of real personal data, but inherits and can amplify biases of the source distribution and may be re-identifiable through membership inference.

Federated learning: model updates rather than raw data leave the device; gradients can leak training examples, so secure aggregation and differential privacy are layered on.

Model drift: performance degrades as the deployment distribution shifts; monitoring for drift is an ethical obligation, not only an engineering one.

Automated decision systems in high-stakes domains (credit, hiring, health, justice) attract stricter scrutiny because errors are consequential and hard to reverse.

Environmental cost: training and inference consume energy and water; sustainability is an emerging ethical consideration in model selection.

Deepfakes and synthetic media: provenance standards (content credentials, watermarking) and disclosure duties are the main mitigation levers.

Dual-use risk: the same capability can serve beneficial and harmful ends; release decisions should weigh uplift to malicious actors against scientific benefit.

Psychological safety: team members can raise concerns without fear of punishment; it is a precondition for surfacing ethical risk early.

Professional codes: ACM Code of Ethics and IEEE Ethically Aligned Design emphasize avoiding harm, honesty, fairness, and respecting privacy and confidentiality.

Conflict of interest: a personal interest that could improperly influence professional judgment; the standard remedy is disclosure plus recusal from the decision.

Whistleblowing channels: internal reporting first, then regulator or designated external body; many jurisdictions protect disclosures made through prescribed channels.

Protected disclosure conditions: the concern is a genuine legal or ethical breach, the disclosure is proportionate, and internal routes were exhausted or demonstrably unsafe.

Retaliation prohibition: adverse action against a good-faith reporter is itself a violation; good faith protects the reporter even if the concern proves unfounded.

Documentation duty: contemporaneous records of concerns raised, responses received, and decisions made are the evidentiary basis for any later claim.

Ethics review boards: independent bodies that review high-risk data projects; their authority depends on a mandate to halt work, not merely to advise.

Tone at the top: leadership behavior, not published policy, is the strongest predictor of whether ethical concerns are reported internally.

Consequentialism: an act is right if its outcomes maximize total welfare; in data work this maps to harm-benefit analysis and cost-benefit review of a model or data release.

Deontology: duties and rules bind regardless of outcome; consent, promise-keeping, and non-deception are constraints that a favorable aggregate result cannot override.

Virtue ethics: evaluates the agent's character (honesty, care, integrity) rather than the act or outcome; used to justify professional dispositions in data practice.

Justice as fairness (Rawls): principles chosen behind a veil of ignorance; supports evaluating a system by how it treats the least advantaged group.

Principlism (Beauchamp and Childress): four mid-level principles - beneficence, non-maleficence, autonomy, justice - commonly extended to data ethics as respect for autonomy, non-harm, beneficence, and fairness.

Moral relativism vs. universalism: relativism ties judgments to culture or context, universalism holds some norms apply across contexts; data ethics codes generally assume a universalist floor.

Fact-value distinction (Hume's is-ought gap): descriptive claims about data do not by themselves entail normative conclusions; a value premise must be supplied explicitly.

Distributive justice criteria: equality (same to all), equity (proportional to need or contribution), and priority to the worst-off yield different allocation rules for data benefits and burdens.

Ethical pluralism: multiple irreducible values (privacy, accuracy, autonomy, welfare) can conflict; resolution requires explicit trade-off justification, not a single metric.

Personal data: any information relating to an identified or identifiable natural person; identifiability can be direct (name) or indirect (a combination of quasi-identifiers).

Data protection principles (GDPR Art. 5): lawfulness, fairness, transparency; purpose limitation; data minimization; accuracy; storage limitation; integrity and confidentiality; accountability.

Legal bases for processing (GDPR Art. 6): consent, contract, legal obligation, vital interests, public task, legitimate interests; special-category data (Art. 9) requires an additional condition.

Data subject rights: access, rectification, erasure, restriction, portability, objection, and rights relating to automated decision-making; erasure is not absolute and yields to legal retention duties.

Consent validity: freely given, specific, informed, unambiguous; silence, pre-ticked boxes, and bundled terms do not constitute valid consent.

Anonymization vs. pseudonymization: pseudonymized data remain personal data because re-identification is possible with additional information held separately; anonymized data fall outside data protection scope.

k-anonymity: each record is indistinguishable from at least k-1 others on quasi-identifiers; it does not protect against attribute disclosure or homogeneity attacks.

Differential privacy: a randomized mechanism satisfies epsilon-differential privacy if outputs for neighboring datasets differ by at most a factor of e^epsilon; smaller epsilon means stronger privacy.

Privacy by design and by default: data protection is embedded in processing from the outset, and the most privacy-protective settings apply without user action.

Statistical parity (demographic parity): the positive decision rate is equal across groups; it ignores base rates and can conflict with predictive accuracy.

Equalized odds: true positive and false positive rates are equal across groups; equal opportunity is the weaker version requiring only equal true positive rates.

Predictive parity (calibration): among those receiving the same score, the outcome rate is equal across groups; calibration within groups and equalized odds generally cannot both hold when base rates differ.

Impossibility result (Chouldechova; Kleinberg et al.): with unequal base rates, calibration, balance for the positive class, and balance for the negative class cannot all be satisfied simultaneously.

Bias sources: historical bias in labels, representation bias in sampling, measurement bias in features, aggregation bias from treating heterogeneous groups as one, and deployment bias from use context.

Proxy variables: features correlated with a protected attribute (postcode, name, school) can transmit disparate impact even when the protected attribute is excluded.

Disparate impact (four-fifths or 80% rule): selection rate for a protected group below 80% of the highest group's rate is a screening signal of adverse impact.

Fairness metric selection rule: the metric must follow from the policy question - allocation decisions favor parity metrics, risk-scoring decisions favor calibration and equalized odds.

Mitigation stages: pre-processing (reweighting, resampling), in-processing (constrained optimization, adversarial debiasing), post-processing (group-specific thresholds).

Transparency types: disclosure of the fact of processing, of the data used, of model logic, and of performance; each answers a different question and satisfies different obligations.

GDPR Art. 13-15: controllers must provide meaningful information about the logic involved in automated decision-making, plus the significance and envisaged consequences.

GDPR Art. 22: data subjects have the right not to be subject to solely automated decisions producing legal or similarly significant effects, subject to exceptions (contract, law, explicit consent).

Interpretability vs. explainability: interpretability is an intrinsic property of simple models; explainability is post-hoc reconstruction of reasons for a complex model's output.

Local vs. global explanation: local explains a single prediction (LIME, SHAP values), global explains overall model behavior (feature importance, surrogate trees).

Explanation quality criteria: faithfulness to the model, stability under small input changes, comprehensibility for the intended audience, and completeness relative to the decision.

Accountability: the obligation to answer for an action; requires an identifiable actor, a documented standard, and a remedy path - transparency alone does not assign responsibility.

Algorithmic impact assessment: structured review of purpose, data, stakeholders, risks, mitigations, and monitoring before deployment, with periodic re-review.

Contestability: affected persons can challenge a decision and obtain a reasoned response; distinct from explanation, which only describes the basis of the decision.

Data governance roles: data owner (accountable for a domain), data steward (quality and policy execution), data custodian (technical safeguards), and the data protection officer (independent oversight).

Data lifecycle stages: acquisition, storage, processing, sharing, retention, archiving, and disposal; controls and ethical duties attach at every stage, not only collection.

CIA triad: confidentiality, integrity, availability; ethics adds accountability and non-maleficence as governance objectives alongside the triad.

Least privilege and need-to-know: access is limited to the minimum necessary for the task and reviewed periodically; role-based access control implements this at scale.

Data classification tiers (public, internal, confidential, restricted) drive handling rules; misclassification is a governance failure even absent a breach.

Retention and deletion: storage limitation requires defined retention periods and verifiable deletion; backups and derived datasets must be included in deletion scope.

Breach notification (GDPR Art. 33-34): notify the supervisory authority within 72 hours of awareness unless unlikely to result in risk; notify subjects when there is high risk.

Security ethics: proportionality of monitoring, no secondary use of security telemetry without a lawful basis, and disclosure of vulnerabilities handled through coordinated disclosure.

GDPR territorial scope (Art. 3): applies to processing in the EU by an establishment, and to non-EU controllers offering goods or services to, or monitoring, EU data subjects.

GDPR penalties: up to EUR 10 million or 2% of global annual turnover for certain obligations; up to EUR 20 million or 4% for core principles and rights, whichever is higher.

Roles: controller determines purposes and means; processor acts on the controller's instructions; a joint-controller arrangement requires an agreed allocation of responsibilities.

DPIA (Art. 35): required for high-risk processing, including large-scale systematic monitoring and large-scale special-category data; consult the supervisory authority if residual risk remains high.

EU AI Act risk tiers: prohibited practices, high-risk systems (conformity assessment, registration, human oversight), limited-risk (transparency duties), minimal-risk (no added obligations).

Sectoral regimes: HIPAA governs protected health information in the US; GLBA covers financial data; COPPA restricts collection from children under 13 in the US.

Cross-border transfers (GDPR Chapter V): adequacy decision, standard contractual clauses, binding corporate rules, or derogations; Schrems II requires a transfer impact assessment.

Legal vs. ethical compliance: lawful processing can still be unethical; regulation sets a floor, and ethical review addresses harms regulation does not reach.

Stakeholder mapping: identify who is affected, the nature and magnitude of impact, power and dependency, and whether affected parties are represented in the decision.

Structured test sequence: facts, stakeholders, options, consequences, duties and rights, virtues, and a final justification that names the trade-off accepted.

Precautionary principle: where harm is plausible and uncertain, take proportionate preventive action rather than waiting for conclusive evidence.

Reversibility test: prefer options whose harms can be undone; irreversible actions (public release, model deployment at scale) demand a higher evidentiary bar.

Harm matrix: rate each option on severity, likelihood, reversibility, and number affected; the product guides but does not replace judgment.

Sunlight (publicity) test: would the decision withstand disclosure to affected users and to peers? Inability to defend it publicly is a warning signal.

Moral residue: after a justified trade-off, some obligation remains unmet; record it and mitigate it rather than declaring the conflict fully resolved.

Case analysis discipline: separate the descriptive question (what does the system do) from the normative question (what should be done), and state the value premise explicitly.

Escalation rule: when a decision exceeds the analyst's authority or the risk is severe and irreversible, refer upward before acting.

Generative models: trained on large corpora and can emit memorized or near-verbatim training content; memorization risk rises with duplication and model capacity.

Training-data provenance: licensing, consent, and copyright status of training corpora are unresolved in many jurisdictions and must be documented for defensibility.

Synthetic data: can reduce exposure of real personal data, but inherits and can amplify biases of the source distribution and may be re-identifiable through membership inference.

Federated learning: model updates rather than raw data leave the device; gradients can leak training examples, so secure aggregation and differential privacy are layered on.

Model drift: performance degrades as the deployment distribution shifts; monitoring for drift is an ethical obligation, not only an engineering one.

Automated decision systems in high-stakes domains (credit, hiring, health, justice) attract stricter scrutiny because errors are consequential and hard to reverse.

Environmental cost: training and inference consume energy and water; sustainability is an emerging ethical consideration in model selection.

Deepfakes and synthetic media: provenance standards (content credentials, watermarking) and disclosure duties are the main mitigation levers.

Dual-use risk: the same capability can serve beneficial and harmful ends; release decisions should weigh uplift to malicious actors against scientific benefit.

Psychological safety: team members can raise concerns without fear of punishment; it is a precondition for surfacing ethical risk early.

Professional codes: ACM Code of Ethics and IEEE Ethically Aligned Design emphasize avoiding harm, honesty, fairness, and respecting privacy and confidentiality.

Conflict of interest: a personal interest that could improperly influence professional judgment; the standard remedy is disclosure plus recusal from the decision.

Whistleblowing channels: internal reporting first, then regulator or designated external body; many jurisdictions protect disclosures made through prescribed channels.

Protected disclosure conditions: the concern is a genuine legal or ethical breach, the disclosure is proportionate, and internal routes were exhausted or demonstrably unsafe.

Retaliation prohibition: adverse action against a good-faith reporter is itself a violation; good faith protects the reporter even if the concern proves unfounded.

Documentation duty: contemporaneous records of concerns raised, responses received, and decisions made are the evidentiary basis for any later claim.

Ethics review boards: independent bodies that review high-risk data projects; their authority depends on a mandate to halt work, not merely to advise.

Tone at the top: leadership behavior, not published policy, is the strongest predictor of whether ethical concerns are reported internally.

Consequentialism: an act is right if its outcomes maximize total welfare; in data work this maps to harm-benefit analysis and cost-benefit review of a model or data release.

Deontology: duties and rules bind regardless of outcome; consent, promise-keeping, and non-deception are constraints that a favorable aggregate result cannot override.

Virtue ethics: evaluates the agent's character (honesty, care, integrity) rather than the act or outcome; used to justify professional dispositions in data practice.

Justice as fairness (Rawls): principles chosen behind a veil of ignorance; supports evaluating a system by how it treats the least advantaged group.

Principlism (Beauchamp and Childress): four mid-level principles - beneficence, non-maleficence, autonomy, justice - commonly extended to data ethics as respect for autonomy, non-harm, beneficence, and fairness.

Moral relativism vs. universalism: relativism ties judgments to culture or context, universalism holds some norms apply across contexts; data ethics codes generally assume a universalist floor.

Fact-value distinction (Hume's is-ought gap): descriptive claims about data do not by themselves entail normative conclusions; a value premise must be supplied explicitly.

Distributive justice criteria: equality (same to all), equity (proportional to need or contribution), and priority to the worst-off yield different allocation rules for data benefits and burdens.

Ethical pluralism: multiple irreducible values (privacy, accuracy, autonomy, welfare) can conflict; resolution requires explicit trade-off justification, not a single metric.

Personal data: any information relating to an identified or identifiable natural person; identifiability can be direct (name) or indirect (a combination of quasi-identifiers).

Data protection principles (GDPR Art. 5): lawfulness, fairness, transparency; purpose limitation; data minimization; accuracy; storage limitation; integrity and confidentiality; accountability.

Legal bases for processing (GDPR Art. 6): consent, contract, legal obligation, vital interests, public task, legitimate interests; special-category data (Art. 9) requires an additional condition.

Data subject rights: access, rectification, erasure, restriction, portability, objection, and rights relating to automated decision-making; erasure is not absolute and yields to legal retention duties.

Consent validity: freely given, specific, informed, unambiguous; silence, pre-ticked boxes, and bundled terms do not constitute valid consent.

Anonymization vs. pseudonymization: pseudonymized data remain personal data because re-identification is possible with additional information held separately; anonymized data fall outside data protection scope.

k-anonymity: each record is indistinguishable from at least k-1 others on quasi-identifiers; it does not protect against attribute disclosure or homogeneity attacks.

Differential privacy: a randomized mechanism satisfies epsilon-differential privacy if outputs for neighboring datasets differ by at most a factor of e^epsilon; smaller epsilon means stronger privacy.

Privacy by design and by default: data protection is embedded in processing from the outset, and the most privacy-protective settings apply without user action.

Statistical parity (demographic parity): the positive decision rate is equal across groups; it ignores base rates and can conflict with predictive accuracy.

Equalized odds: true positive and false positive rates are equal across groups; equal opportunity is the weaker version requiring only equal true positive rates.

Predictive parity (calibration): among those receiving the same score, the outcome rate is equal across groups; calibration within groups and equalized odds generally cannot both hold when base rates differ.

Impossibility result (Chouldechova; Kleinberg et al.): with unequal base rates, calibration, balance for the positive class, and balance for the negative class cannot all be satisfied simultaneously.

Bias sources: historical bias in labels, representation bias in sampling, measurement bias in features, aggregation bias from treating heterogeneous groups as one, and deployment bias from use context.

Proxy variables: features correlated with a protected attribute (postcode, name, school) can transmit disparate impact even when the protected attribute is excluded.

Disparate impact (four-fifths or 80% rule): selection rate for a protected group below 80% of the highest group's rate is a screening signal of adverse impact.

Fairness metric selection rule: the metric must follow from the policy question - allocation decisions favor parity metrics, risk-scoring decisions favor calibration and equalized odds.

Mitigation stages: pre-processing (reweighting, resampling), in-processing (constrained optimization, adversarial debiasing), post-processing (group-specific thresholds).

Transparency types: disclosure of the fact of processing, of the data used, of model logic, and of performance; each answers a different question and satisfies different obligations.

GDPR Art. 13-15: controllers must provide meaningful information about the logic involved in automated decision-making, plus the significance and envisaged consequences.

GDPR Art. 22: data subjects have the right not to be subject to solely automated decisions producing legal or similarly significant effects, subject to exceptions (contract, law, explicit consent).

Interpretability vs. explainability: interpretability is an intrinsic property of simple models; explainability is post-hoc reconstruction of reasons for a complex model's output.

Local vs. global explanation: local explains a single prediction (LIME, SHAP values), global explains overall model behavior (feature importance, surrogate trees).

Explanation quality criteria: faithfulness to the model, stability under small input changes, comprehensibility for the intended audience, and completeness relative to the decision.

Accountability: the obligation to answer for an action; requires an identifiable actor, a documented standard, and a remedy path - transparency alone does not assign responsibility.

Algorithmic impact assessment: structured review of purpose, data, stakeholders, risks, mitigations, and monitoring before deployment, with periodic re-review.

Contestability: affected persons can challenge a decision and obtain a reasoned response; distinct from explanation, which only describes the basis of the decision.

Data governance roles: data owner (accountable for a domain), data steward (quality and policy execution), data custodian (technical safeguards), and the data protection officer (independent oversight).

Data lifecycle stages: acquisition, storage, processing, sharing, retention, archiving, and disposal; controls and ethical duties attach at every stage, not only collection.

CIA triad: confidentiality, integrity, availability; ethics adds accountability and non-maleficence as governance objectives alongside the triad.

Least privilege and need-to-know: access is limited to the minimum necessary for the task and reviewed periodically; role-based access control implements this at scale.

Data classification tiers (public, internal, confidential, restricted) drive handling rules; misclassification is a governance failure even absent a breach.

Retention and deletion: storage limitation requires defined retention periods and verifiable deletion; backups and derived datasets must be included in deletion scope.

Breach notification (GDPR Art. 33-34): notify the supervisory authority within 72 hours of awareness unless unlikely to result in risk; notify subjects when there is high risk.

Security ethics: proportionality of monitoring, no secondary use of security telemetry without a lawful basis, and disclosure of vulnerabilities handled through coordinated disclosure.

GDPR territorial scope (Art. 3): applies to processing in the EU by an establishment, and to non-EU controllers offering goods or services to, or monitoring, EU data subjects.

GDPR penalties: up to EUR 10 million or 2% of global annual turnover for certain obligations; up to EUR 20 million or 4% for core principles and rights, whichever is higher.

Roles: controller determines purposes and means; processor acts on the controller's instructions; a joint-controller arrangement requires an agreed allocation of responsibilities.

DPIA (Art. 35): required for high-risk processing, including large-scale systematic monitoring and large-scale special-category data; consult the supervisory authority if residual risk remains high.

EU AI Act risk tiers: prohibited practices, high-risk systems (conformity assessment, registration, human oversight), limited-risk (transparency duties), minimal-risk (no added obligations).

Sectoral regimes: HIPAA governs protected health information in the US; GLBA covers financial data; COPPA restricts collection from children under 13 in the US.

Cross-border transfers (GDPR Chapter V): adequacy decision, standard contractual clauses, binding corporate rules, or derogations; Schrems II requires a transfer impact assessment.

Legal vs. ethical compliance: lawful processing can still be unethical; regulation sets a floor, and ethical review addresses harms regulation does not reach.

Stakeholder mapping: identify who is affected, the nature and magnitude of impact, power and dependency, and whether affected parties are represented in the decision.

Structured test sequence: facts, stakeholders, options, consequences, duties and rights, virtues, and a final justification that names the trade-off accepted.

Precautionary principle: where harm is plausible and uncertain, take proportionate preventive action rather than waiting for conclusive evidence.

Reversibility test: prefer options whose harms can be undone; irreversible actions (public release, model deployment at scale) demand a higher evidentiary bar.

Harm matrix: rate each option on severity, likelihood, reversibility, and number affected; the product guides but does not replace judgment.

Sunlight (publicity) test: would the decision withstand disclosure to affected users and to peers? Inability to defend it publicly is a warning signal.

Moral residue: after a justified trade-off, some obligation remains unmet; record it and mitigate it rather than declaring the conflict fully resolved.

Case analysis discipline: separate the descriptive question (what does the system do) from the normative question (what should be done), and state the value premise explicitly.

Escalation rule: when a decision exceeds the analyst's authority or the risk is severe and irreversible, refer upward before acting.

Generative models: trained on large corpora and can emit memorized or near-verbatim training content; memorization risk rises with duplication and model capacity.

Training-data provenance: licensing, consent, and copyright status of training corpora are unresolved in many jurisdictions and must be documented for defensibility.

Synthetic data: can reduce exposure of real personal data, but inherits and can amplify biases of the source distribution and may be re-identifiable through membership inference.

Federated learning: model updates rather than raw data leave the device; gradients can leak training examples, so secure aggregation and differential privacy are layered on.

Model drift: performance degrades as the deployment distribution shifts; monitoring for drift is an ethical obligation, not only an engineering one.

Automated decision systems in high-stakes domains (credit, hiring, health, justice) attract stricter scrutiny because errors are consequential and hard to reverse.

Environmental cost: training and inference consume energy and water; sustainability is an emerging ethical consideration in model selection.

Deepfakes and synthetic media: provenance standards (content credentials, watermarking) and disclosure duties are the main mitigation levers.

Dual-use risk: the same capability can serve beneficial and harmful ends; release decisions should weigh uplift to malicious actors against scientific benefit.

Psychological safety: team members can raise concerns without fear of punishment; it is a precondition for surfacing ethical risk early.

Professional codes: ACM Code of Ethics and IEEE Ethically Aligned Design emphasize avoiding harm, honesty, fairness, and respecting privacy and confidentiality.

Conflict of interest: a personal interest that could improperly influence professional judgment; the standard remedy is disclosure plus recusal from the decision.

Whistleblowing channels: internal reporting first, then regulator or designated external body; many jurisdictions protect disclosures made through prescribed channels.

Protected disclosure conditions: the concern is a genuine legal or ethical breach, the disclosure is proportionate, and internal routes were exhausted or demonstrably unsafe.

Retaliation prohibition: adverse action against a good-faith reporter is itself a violation; good faith protects the reporter even if the concern proves unfounded.

Documentation duty: contemporaneous records of concerns raised, responses received, and decisions made are the evidentiary basis for any later claim.

Ethics review boards: independent bodies that review high-risk data projects; their authority depends on a mandate to halt work, not merely to advise.

Tone at the top: leadership behavior, not published policy, is the strongest predictor of whether ethical concerns are reported internally.