CDEI flashcards
CDEI · Build recall with a focused review session.
32 cards. Read the term and the definition here. The deck at the top of this page is optional.
- Data ethics
- The branch of ethics concerned with the moral principles and values governing the collection, processing, sharing, and use of data, distinct from legal compliance and technical governance.
- Collingridge dilemma
- The problem that a technology's impacts are unknown early on, but once known the technology is entrenched and difficult to change, complicating timely ethical regulation.
- Ethical lag
- The gap between the rapid pace of technological change and the slower development of ethical norms, laws, and policies needed to govern it responsibly.
- Stakeholder analysis
- A structured method for identifying all parties affected by a data practice and assessing the harms and benefits to each before deciding whether to proceed.
- Consequentialism
- An ethical framework judging actions by their outcomes, holding that the right action produces the greatest overall balance of benefits over harms.
- Deontology
- An ethical framework holding that certain actions are inherently right or wrong regardless of consequences, and that people must be treated as ends, not merely means.
- Virtue ethics
- An ethical framework focused on the character traits of the actor, such as honesty, integrity, and care, rather than on rules or consequences.
- Veil of ignorance
- A Rawlsian thought experiment in which principles of justice are chosen without knowing one's own position, ensuring fairness to the least advantaged.
- Value alignment
- The design goal of ensuring that an AI system's behavior and objectives reflect the intended values and preferences of its users and society.
- k-anonymity
- A privacy property requiring that each record be indistinguishable from at least k-1 other records on quasi-identifiers, reducing but not eliminating re-identification risk.
- Differential privacy
- A formal privacy guarantee in which a randomized mechanism's output distributions on neighboring datasets differ by at most a factor of e^ε, with smaller ε giving stronger privacy.
- Data minimization
- The principle that only personal data necessary for a specified purpose should be collected and that it should be retained only as long as needed.
- Informed consent
- Authorization that is specific, freely given, informed about purpose and risks, and revocable; bundled or coerced consent is ethically invalid.
- Re-identification risk
- The chance that de-identified data can be linked to individuals by combining it with auxiliary datasets, as demonstrated in the Netflix Prize and AOL cases.
- Privacy by design
- An approach that embeds privacy protections into system architecture and default settings from the outset rather than adding them after deployment.
- Historical bias
- Bias that arises when training data reflects past inequities, causing a model to reproduce and amplify those inequities in its predictions.
- Measurement bias
- Bias occurring when a proxy variable is used for a construct and measures that construct differently across groups, distorting model behavior.
- Statistical parity
- A fairness criterion requiring equal positive prediction rates across groups, regardless of differences in underlying base rates or error rates.
- Equalized odds
- A fairness criterion requiring equal true positive and false positive rates across groups; equal opportunity requires only equal true positive rates.
- Disparate impact
- A facially neutral practice that disproportionately harms a protected group; the 80% rule flags selection rate ratios below 0.8.
- Data governance
- The system of decision rights, policies, and accountability mechanisms that governs data assets across their lifecycle within an organization.
- Data stewardship
- The operational responsibility for data quality, access, and appropriate use, typically assigned to designated data stewards within a governance framework.
- Data provenance
- Documentation of the origin, lineage, and transformations of data, enabling auditability, reproducibility, and trust in analytical results.
- Least privilege
- An access control principle granting users only the minimum permissions needed for their tasks, reducing the risk of misuse or breach.
- Data protection impact assessment
- A structured assessment required for high-risk processing under GDPR that identifies privacy risks and mitigations before a project begins.
- Explainability
- The ability to provide understandable reasons for a specific model output or for a model's overall behavior, supporting trust and accountability.
- SHAP
- A post-hoc explanation method based on Shapley values that attributes a model's prediction to individual features in a theoretically grounded way.
- Accountability
- The obligation to answer for decisions and their consequences, requiring identifiable responsible parties and mechanisms for remedy.
- Model card
- A document reporting a model's intended use, performance across groups, limitations, and ethical considerations to support informed deployment decisions.
- GDPR
- The EU General Data Protection Regulation, establishing principles, lawful bases, data subject rights, and accountability obligations for processing personal data.
- CCPA/CPRA
- California privacy laws granting consumers rights to know, delete, correct, opt out of sale or sharing, and limit use of sensitive personal information.
- EU AI Act
- A risk-based regulation classifying AI systems into prohibited, high-risk, limited-risk, and minimal-risk tiers, with obligations increasing with risk level.