CDEF key terms

The branch of applied ethics that examines how data is collected, processed, shared, and used, and how those practices affect people, groups, and society. It goes beyond mere legal compliance to ask what is right.

An ethical framework judging an action by its outcomes, such as maximizing overall welfare. In data work it supports cost-benefit analysis of harms and benefits but can justify sacrificing minorities for aggregate gains.

An ethical framework holding that certain duties and rules bind us regardless of consequences, such as respecting persons and keeping promises. It grounds absolute prohibitions on deception and treating people merely as data sources.

An approach focused on the character traits of the actor rather than on rules or outcomes. It asks what an honest, careful, and just data professional would do in a given situation.

A systematic framework, such as consequentialism, deontology, or virtue ethics, used to reason about right and wrong. It gives structure to justifications so decisions rest on principle rather than intuition alone.

A person's voluntary agreement to data processing after receiving clear information about purpose, scope, risks, and alternatives. Valid consent requires capacity, comprehension, and the genuine freedom to refuse without penalty.

The principle that personal data collected for specified purposes must not be reused for incompatible new purposes without a fresh legal basis. It prevents function creep where data quietly migrates to unrelated uses.

Collecting and retaining only the personal data that is adequate, relevant, and necessary for a stated purpose. It reduces exposure to breach, misuse, and overreach by shrinking the data footprint.

The principle that personal data should be kept only as long as needed for its purpose, then deleted or anonymized. It counters indefinite retention driven by speculative future value.

The entitlements of individuals over their personal data, typically including access, correction, erasure, restriction, portability, and objection. They shift control toward the person the data describes.

An individual's right to have personal data deleted in defined circumstances, such as when it is no longer necessary or consent is withdrawn. Also called the right to be forgotten, it is not absolute.

Processing that irreversibly removes identifying characteristics so a person can no longer be singled out, linked, or inferred. True anonymization is hard; weak de-identification leaves re-identification risk.

Replacing identifiers with codes so data cannot be attributed to a person without additional separately held information. It reduces exposure but the data remains personal because re-linking is possible.

Recovering the identity of individuals from supposedly de-identified data by linking it with other datasets or exploiting unique attribute combinations. It shows that removing names alone rarely protects privacy.

Systematic and unfair error or disadvantage produced by an algorithm's design, training data, or deployment context. It can disadvantage groups even when protected attributes are never used directly.

A feature that stands in for a protected characteristic, such as postcode for race or first name for gender. Models using proxies can discriminate while appearing to ignore protected attributes.

A neutral-seeming policy or model that produces disproportionately adverse outcomes for a protected group, regardless of intent. It is a key legal and ethical test for indirect discrimination.

A personal characteristic, such as race, sex, age, religion, or disability, that laws shield from discriminatory treatment. Data work must handle these fields with heightened care and justification.

A quantified criterion for comparing model outcomes across groups, such as demographic parity, equal opportunity, or equalized odds. Metrics conflict mathematically, so choosing one is an ethical and contextual judgment.

The degree to which a person can understand why a system produced a given output. It ranges from inherently interpretable models to post-hoc explanations of complex ones tailored to the audience.

The property of a model whose internal logic is directly understandable by humans, such as a small decision tree or linear model. It trades some predictive power for clarity and contestability.

Openness about what data is used, how systems work, and who is accountable, appropriate to the audience. It enables scrutiny and trust but is not the same as full disclosure of proprietary detail.

The obligation to answer for decisions and outcomes and to remedy harm. In data work it requires named owners, documented decisions, audit trails, and real consequences when things go wrong.

A structured review of a proposed or live system's effects on rights, safety, and groups before and after deployment. It documents risks, mitigations, and sign-off, similar to a data protection impact assessment.

The responsible management of data assets across their lifecycle to ensure quality, appropriate access, and ethical use. Stewards act as custodians for others rather than owners of the data.

The policies, roles, and processes that define who may access, change, and use data, and under what conditions. It turns ethical intent into enforceable operating rules.

Technical and organizational rules limiting who can view or use a resource, based on roles, need-to-know, or attributes. It enforces least privilege and limits the blast radius of misuse.

Granting each user or system only the minimum access needed for a task, for the shortest necessary time. It reduces both accidental error and the damage from compromised accounts.

A documented evaluation of privacy risks before processing likely to be high-risk, such as large-scale profiling. It identifies harms, safeguards, and whether processing should proceed at all.

The European Union's comprehensive data protection law governing processing of personal data. It sets principles, legal bases, data subject rights, breach notification duties, and heavy fines for violations.

The lawful ground that authorizes handling personal data, such as consent, contract, legal obligation, or legitimate interest. Without one, processing is unlawful regardless of good intentions.

The duty to report certain security incidents to regulators and affected individuals within set timeframes. It reflects the principle that people deserve to know when their data is exposed.

A repeatable process for resolving dilemmas, typically identifying stakeholders, facts, options, principles, and consequences before deciding and documenting the rationale. It makes reasoning reviewable and consistent.

The discomfort felt when a professional knows the right action but is constrained from taking it by organizational pressure or rules. Unaddressed, it drives disengagement and silence.

Reporting wrongdoing, such as unlawful data use or concealment of harm, to an internal or external authority. Ethical guidance favors exhausting internal channels first when they are credible and safe.

A report of wrongdoing that the law shields from retaliation, provided it meets defined conditions such as good faith and proper channels. It balances loyalty to an employer against public interest.

Consistently acting on ethical commitments, such as accuracy and honesty, even under commercial pressure. It includes refusing to misrepresent data or hide known model limitations.

A short standardized document describing a model's intended use, training data, performance across groups, and limitations. It supports informed adoption and external scrutiny.

Artificially generated records that mimic the statistical patterns of real data without describing actual individuals. It can reduce privacy risk but may inherit or amplify biases from its source.