Fairness, bias, transparency, accountability and the wider social effects of AI systems.
56 terms
AI accountability means that identifiable people and organizations answer for how an AI system is built and used, and must explain and remedy its harms.
AI anthropomorphism is the attribution of human traits such as feelings, intentions or understanding to AI systems, a central concern in AI ethics.
An AI audit is a structured, evidence-based examination of an AI system that tests its fairness, accuracy, transparency and compliance against defined criteria.
AI disclosure is the practice of telling people when they interact with an AI system or see content that AI generated or substantially altered.
AI environmental impact covers the energy, carbon emissions, water and hardware resources consumed across the life cycle of artificial intelligence systems.
AI ethics is the field that studies and governs fairness, transparency, accountability, privacy and social impact in the design and use of AI systems.
An AI ethics board is a standing body that reviews how an organization builds and uses AI systems and advises on fairness, risk and accountability.
The AI Incident Database (AIID) is an open archive of documented real-world harms caused by AI systems, used to learn from failures and prevent repeats.
AI job displacement is the loss or reshaping of jobs when AI systems take over tasks workers once did, raising ethical questions of fairness and transition.
AI moral status is the question in AI ethics of whether an AI system could have interests that matter morally for its own sake, such as…
AI transparency is the practice of making clear how an AI system works, what data shapes it, where it is used and what its limits are.
Algorithmic fairness is the study and practice of ensuring that automated decisions do not systematically disadvantage people based on protected traits.
An algorithmic impact assessment is a structured review of how an automated decision system may affect people, rights and communities before and during use.
Allocative harm occurs when an AI system unfairly withholds opportunities or resources, such as loans, jobs or housing, from certain groups of people.
Automation bias is the tendency to over-trust automated or AI output, accepting wrong suggestions and missing problems the system fails to flag.
Bias amplification occurs when a machine learning model exaggerates skews in its training data, producing outputs more unequal than the data itself.
Bias mitigation, or debiasing, covers the techniques that reduce unfair, group-dependent errors in machine learning data, models and decisions.
A black-box model is an AI system whose internal reasoning cannot be inspected or understood, raising transparency and accountability concerns.
Contestability is the ability of people affected by an AI decision to challenge it, have it reviewed by a human and obtain correction where it is…
A counterfactual explanation tells a person the smallest realistic change to their inputs that would have flipped an AI model's decision.
Counterfactual fairness requires that a model's decision for a person stay the same had their protected attribute been different, under a causal model.
Datasheets for datasets are structured documents that record how a dataset was built, what it contains and how it should and should not be used.
Demographic parity is a group fairness criterion requiring that an AI model gives positive outcomes at equal rates across protected groups.
Disparate impact is unintended discrimination in which a neutral rule or AI model produces substantially worse outcomes for a protected group.
The ELIZA effect is the tendency to attribute understanding, feelings or intent to a computer program because it produces humanlike text.
Equal opportunity is a group fairness criterion requiring a classifier to have equal true positive rates for qualified people across protected groups.
Equalized odds is a group fairness criterion requiring a classifier to have equal true positive and false positive rates across protected groups.
Ethics washing is the practice of displaying AI ethics principles, boards or pledges to look responsible without changing how AI systems are built or used.
A fairness metric is a quantitative measure that tests whether an AI model's predictions or errors differ unjustifiably across groups or individuals.
Fairness through unawareness is the approach of excluding protected attributes from a model, which rarely prevents bias because proxies remain in the data.
The fairness-accuracy trade-off describes how enforcing fairness constraints on a machine learning model can reduce its measured predictive accuracy.
The four-fifths rule flags possible adverse impact when one group's selection rate falls below 80 percent of the most favored group's rate.
Gender bias in AI is systematic unequal treatment or portrayal of people by gender in AI outputs, caused by skewed data, design choices and deployment.
Ghost work is the hidden human labor of labeling, moderating and checking data that AI systems depend on, often done by low-paid remote contractors.
Grad-CAM is a post-hoc explanation method that highlights which image regions most influenced a convolutional network's prediction for a chosen class.
Green AI is research and practice that treats energy use, carbon emissions and computational efficiency as core criteria when building AI systems.
Group fairness is a family of criteria requiring a machine learning model to treat demographic groups comparably in its decisions or error rates.
Historical bias is unfairness a machine learning model inherits from data that accurately records a world shaped by past discrimination and inequality.
Human-AI collaboration is the joint work of people and AI systems in which each contributes distinct strengths and humans keep meaningful accountability.
Human-centered AI (HCAI) designs AI systems around human needs, control and well-being, so that people stay informed, empowered and accountable.
Individual fairness is the AI ethics principle that a model should treat similar individuals similarly, judged by a task-specific similarity metric.
Integrated gradients is an attribution method that explains a neural network prediction by summing gradients along a path from a baseline to the input.
An interpretable model is a machine learning model whose structure lets people directly see how inputs lead to each prediction, without separate explainers.
LIME explains a single prediction of any black-box model by fitting a simple, interpretable surrogate model to perturbed samples around that one input.
Machine ethics is the field that designs AI systems able to represent moral norms and reason about them when choosing actions with consequences for people.
Measurement bias in AI arises when features or labels are measured, recorded or defined differently across groups, distorting what a model learns.
A partial dependence plot shows how a model's average prediction changes as one or two features vary, supporting transparency reviews in AI ethics.
A post-hoc explanation describes why an already trained AI model produced a prediction, without changing the model or how it reached its decision.
Predictive parity is a group fairness criterion requiring that a positive prediction be equally reliable, with equal precision, for every protected group.
A protected attribute is a personal characteristic, such as race, sex or age, on which an AI system must not base unfair or unlawful differences in…