Appendix A: Glossary of Key Terms

This glossary defines the principal terms, concepts, legal frameworks, and technical vocabulary used throughout the textbook. Terms appear in alphabetical order. Where a term is closely related to another entry, a cross-reference is provided in parentheses. Definitions are written for a business professional audience and aim to be precise without requiring a computer science or philosophy background.

A

Accountability The obligation of an individual, team, or organization to answer for decisions made by or with an AI system, and to bear consequences when those decisions cause harm. Accountability requires clear chains of responsibility — someone must own the outcome, not just the process. See also: Governance, Principal-agent problem, Oversight.

Accuracy A statistical measure of how often a model's predictions match actual outcomes, calculated as the number of correct predictions divided by total predictions. Accuracy can be misleading in imbalanced datasets (e.g., if 95% of loan applicants are approved, a model that always predicts approval achieves 95% accuracy while being useless). See also: True positive/false positive, Calibration.

Adverse Action Notice A legally required disclosure, under U.S. law (e.g., the Equal Credit Opportunity Act and Fair Credit Reporting Act), informing a consumer that credit, employment, insurance, or housing has been denied or limited, and providing the principal reasons. Algorithmic systems that automate these decisions must still produce legally adequate adverse action notices. See also: ECOA, Explainability, Right to explanation.

Adversarial Attack A deliberate manipulation of an AI system's inputs — often imperceptible to humans — designed to cause incorrect outputs. Adversarial attacks exploit the mathematical vulnerabilities of machine learning models and represent a significant security concern in high-stakes deployments. See also: Adversarial example, Data poisoning, Red-teaming.

Adversarial Example A specific input, often visually or textually indistinguishable from a normal input, that causes a trained model to make a confident but incorrect prediction. A classic example is adding carefully calibrated pixel noise to an image of a panda that causes the model to classify it as a gibbon with high confidence. See also: Adversarial attack.

Aggregation Bias A form of bias arising when a model is trained on pooled data from heterogeneous groups whose underlying patterns differ significantly. The model learns an average that fits no subgroup well. In medicine, models trained on combined patient populations often underperform for minority groups whose physiology or disease presentation differs from the majority. See also: Representation bias, Distribution shift.

AI Act (EU) See: EU AI Act.

AI Ethics The interdisciplinary field that applies moral reasoning, social science, legal analysis, and technical expertise to questions raised by the design, deployment, and governance of artificial intelligence systems. AI ethics addresses questions of fairness, accountability, transparency, privacy, safety, and the distribution of AI's benefits and harms. See also: Responsible AI, Governance.

AI Governance The policies, procedures, standards, institutional structures, legal frameworks, and technical controls through which organizations and governments guide the development and deployment of AI systems. Effective AI governance connects organizational values to operational AI practices. See also: Governance, NIST AI RMF, EU AI Act.

AI Safety A research and practice area concerned with ensuring that AI systems behave as intended, remain under meaningful human control, and do not cause catastrophic or irreversible harm. AI safety encompasses both near-term concerns (system reliability, misuse, bias) and longer-term concerns about advanced AI systems. See also: Oversight, Human-in-the-loop, Autonomous weapons.

Algorithmic Accountability The principle that those who design, deploy, and profit from algorithmic systems should be answerable for those systems' outcomes, including harmful or discriminatory ones. Algorithmic accountability requires audit mechanisms, transparency, and enforceable remedies. See also: Accountability, Algorithmic transparency, Auditing AI systems.

Algorithmic Bias Systematic and unfair discrimination produced by an algorithm, often reflecting or amplifying human biases embedded in training data, problem framing, or model design. Algorithmic bias can arise even when protected characteristics such as race or gender are not explicitly included as features. See also: Bias (discriminatory), Proxy variable, Historical bias.

Algorithmic Decision-Making The use of automated computational processes to make or support decisions that affect individuals or groups. When consequential (affecting credit, employment, healthcare, or criminal justice), algorithmic decision-making raises significant ethical, legal, and accountability questions. See also: Algorithmic accountability, High-risk AI.

Algorithmic Impact Assessment (AIA) A structured evaluation — conducted before or after deployment — of the potential harms, benefits, and fairness implications of an algorithmic system. Modeled on environmental impact assessments, AIAs are increasingly required or recommended by regulators and governance frameworks. See also: Impact assessment, Auditing AI systems.

Algorithmic Redlining The digital-era analog of historical geographic redlining, in which algorithmic systems systematically deny services, opportunities, or favorable terms to residents of certain neighborhoods or to people sharing demographic characteristics. The discrimination is encoded in proxy variables rather than explicit geographic boundaries. See also: Redlining, Proxy variable, Disparate impact.

Algorithmic Transparency The degree to which information about how an algorithmic system works — its data sources, logic, decision criteria, and performance — is accessible to relevant stakeholders. Transparency is necessary but not sufficient for accountability. See also: Explainability, Transparency, Black box.

Artificial Intelligence (AI) The field of computer science concerned with creating systems that perform tasks requiring capabilities typically associated with human intelligence — such as perception, reasoning, learning, planning, and language understanding. In practice, most deployed AI is narrow AI, optimized for specific tasks rather than general reasoning. See also: Machine learning, Deep learning, Foundation model.

Attention Mechanism A neural network component that allows a model to weight different parts of an input sequence (e.g., different words in a sentence) when generating an output, enabling the model to "focus" on the most relevant information. Attention mechanisms are foundational to transformer architectures and large language models. See also: Large language model, Neural network.

Automation Bias The tendency of humans to over-rely on automated recommendations and to underweight their own judgment or contradictory information when an AI system is involved. Automation bias reduces the effectiveness of human-in-the-loop oversight and is especially dangerous in high-stakes medical or legal applications. See also: Human-in-the-loop, Oversight.

Autonomous Weapons (LAWS) See: LAWS (Lethal Autonomous Weapons Systems).

B

Backdoor Attack A type of adversarial attack in which malicious actors embed hidden triggers into a model during training. The model performs normally on standard inputs but produces attacker-specified outputs when the trigger is present in a deployment-time input. Also called a trojan attack. See also: Adversarial attack, Data poisoning.

Base Rate The underlying frequency of an outcome in a population before any model prediction is applied. Ignoring base rates leads to misinterpretation of model performance: a model with 99% accuracy detecting a disease with 0.1% prevalence will still produce more false positives than true positives. See also: Accuracy, True positive/false positive, Calibration.

Benchmark A standardized dataset or evaluation task used to measure and compare the performance of AI models. Benchmarks enable progress tracking but can also create distorted incentives if optimizing for benchmark performance diverges from real-world usefulness or safety. See also: Accuracy, Model card.

Bias (Statistical) A systematic error in a statistical estimate or model output that consistently deviates from the true value in one direction. Statistical bias is a technical property distinct from discriminatory bias, though the two often interact. See also: Bias (discriminatory), Measurement bias.

Bias (Discriminatory) A systematic pattern in AI outputs that disadvantages individuals or groups on the basis of characteristics such as race, gender, age, disability, or national origin. Discriminatory bias can arise from biased training data, problem framing, feature selection, or optimization criteria. See also: Algorithmic bias, Disparate impact, Historical bias.

Black Box An AI system whose internal reasoning process is opaque or inaccessible to human inspection. Deep neural networks are archetypal black boxes: they may produce accurate outputs but without an interpretable explanation of why. See also: Explainability, Interpretability, Explainable AI.

Brussels Effect The phenomenon by which stringent EU regulations — particularly in data protection and AI — effectively set global standards because multinational companies find it impractical to maintain separate product lines for different markets. The GDPR and EU AI Act are primary examples. See also: GDPR, EU AI Act, Governance.

C

Calibration The degree to which a model's predicted probabilities match actual outcome frequencies. A well-calibrated model that assigns 70% probability to an event is correct approximately 70% of the time. Poor calibration can be a form of bias, particularly across demographic groups. See also: Accuracy, Equalized odds.

Cambridge Analytica A political consulting firm that harvested personal data from up to 87 million Facebook users without explicit consent and used it to build psychographic profiles for micro-targeted political advertising in the 2016 U.S. election and Brexit referendum. The scandal catalyzed global debate about data privacy, platform accountability, and the weaponization of AI in democratic processes. See also: Surveillance capitalism, GDPR, Democratic processes.

Capabilities Approach An ethical framework developed by Amartya Sen and Martha Nussbaum that evaluates well-being by reference to what people are actually able to do and be, rather than by resources or preferences alone. Applied to AI ethics, it asks whether AI systems expand or constrain people's real freedoms and capabilities. See also: Consequentialism, Justice.

Care Ethics An ethical framework, associated with Carol Gilligan and Nel Noddings, that centers relationships, context, and responsiveness to particular needs rather than abstract universal rules. Care ethics challenges purely procedural or rule-based approaches to AI fairness by emphasizing the importance of context, relationship, and attentiveness to vulnerability. See also: Virtue ethics, Consequentialism, Deontology.

Categorical Imperative The central principle of Kant's deontological ethics, which holds that one should act only according to rules one could universalize — that is, apply consistently to all rational beings. Applied to AI, the categorical imperative would prohibit practices (like mass surveillance without consent) that could not be justified as universal principles. See also: Deontology, Contractualism.

CCPA/CPRA (California Consumer Privacy Act / California Privacy Rights Act) California's comprehensive consumer privacy laws, effective 2020 (CCPA) and 2023 (CPRA), granting California residents rights to know what personal data is collected about them, to delete it, to opt out of its sale, and to limit the use of sensitive personal data. Among the most significant U.S. state privacy laws. See also: GDPR, Privacy by design, Data subject.

Civil Rights Legal and constitutional protections against discrimination on the basis of protected characteristics such as race, color, national origin, sex, disability, and religion. U.S. civil rights laws — including Title VII, the Fair Housing Act, and the ECOA — apply to algorithmic systems used in employment, lending, and housing. See also: Title VII, Fair Housing Act, ECOA, Disparate impact.

Classification A machine learning task in which a model assigns an input to one of a predefined set of categories (e.g., loan approved/denied, email spam/not spam, image contains face/no face). Many high-stakes algorithmic decisions are framed as classification problems. See also: Machine learning, Logistic regression.

Clearview AI A facial recognition company that scraped billions of images from social media and public websites to build a face identification database, marketed primarily to law enforcement. Clearview AI raised profound privacy and civil liberties concerns and has faced regulatory action in multiple countries. See also: Facial recognition, Biometrics, Surveillance capitalism.

COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) A commercially developed algorithmic risk assessment tool used in the U.S. criminal justice system to predict recidivism risk. A 2016 ProPublica investigation found that COMPAS incorrectly labeled Black defendants as high-risk at roughly twice the rate it did white defendants, sparking a landmark debate about fairness in criminal justice algorithms. See also: Recidivism, Algorithmic bias, Risk assessment (criminal justice).

Confucian Ethics An ethical tradition rooted in the teachings of Confucius, emphasizing relational duties, social harmony, benevolence (ren), righteousness (yi), ritual propriety (li), and the cultivation of virtue. Confucian ethics offers a relational and communitarian counterpoint to Western individualist ethics in global AI governance debates. See also: Ubuntu, Care ethics, Virtue ethics.

Consequentialism An ethical theory holding that the morality of an action is determined solely by its consequences — specifically, whether it maximizes good outcomes (e.g., well-being, utility) and minimizes harm. Utilitarian approaches to AI governance (maximizing aggregate benefit) are a form of consequentialism. See also: Deontology, Virtue ethics, Capabilities approach.

Contractualism An ethical theory holding that moral principles are those that no one could reasonably reject as a basis for mutual governance. T.M. Scanlon's version emphasizes what principles rational agents would agree to under fair conditions. Contractualism supports procedural fairness and consent-based approaches to AI governance. See also: Veil of ignorance, Categorical imperative.

Contrastive Explanation An explanation that answers the question "why P rather than Q?" — explaining not just why a particular decision was reached but why an alternative outcome did not occur. Contrastive explanations are often more useful to affected individuals than simple feature importance scores. See also: Counterfactual explanation, Explainability, XAI.

Counterfactual Explanation An explanation that describes the minimal change to an input that would have produced a different output (e.g., "Your loan would have been approved if your income were $5,000 higher"). Counterfactual explanations are practically useful for affected individuals seeking to understand and challenge automated decisions. See also: Contrastive explanation, Explainability, Adverse action notice.

Counterfactual Fairness A formal fairness criterion requiring that a prediction for an individual would remain the same in a counterfactual world where their race, gender, or other protected attribute were different, holding all causally downstream variables constant. It operationalizes the intuition that protected characteristics should not cause unfavorable outcomes. See also: Individual fairness, Demographic parity, Causal inference.

D

Data Annotation The process of labeling raw data (images, text, audio, etc.) with metadata or classifications to create training data for machine learning models. Annotation work is often performed by low-paid contract workers, raising labor and accuracy concerns, and the labels themselves can embed human biases. See also: Ghost work, Training data, Label bias.

Data Augmentation Techniques that artificially expand the size and diversity of training datasets by creating modified versions of existing data (e.g., rotating images, adding noise, paraphrasing text). Data augmentation can help reduce representation bias by increasing coverage of underrepresented cases. See also: Training data, Representation bias.

Data Colonialism A theoretical framework describing how data extraction — particularly from communities in the Global South — replicates dynamics of historical colonialism: collecting resources from marginalized populations for the benefit of concentrated corporate power, with little return to the source communities. See also: Surveillance capitalism, Global AI governance, Representation bias.

Data Minimization The privacy principle requiring that only data strictly necessary for a specified purpose be collected and retained. Data minimization limits exposure to breach, misuse, and surveillance, and is a core requirement of the GDPR. See also: Privacy by design, Purpose limitation, GDPR.

Data Poisoning An attack in which malicious actors introduce corrupted or adversarial data into a model's training set to degrade performance or embed backdoors. Data poisoning is particularly concerning for models trained on data sourced from the public internet. See also: Adversarial attack, Backdoor attack, Cybersecurity.

Data Subject Under the GDPR and similar laws, the identified or identifiable natural person to whom personal data relates. Data subjects have rights including access, correction, deletion, and in some contexts the right to object to automated decision-making. See also: GDPR, Privacy by design, Right to explanation.

Deep Learning A subfield of machine learning using artificial neural networks with many layers (hence "deep") to learn hierarchical representations from data. Deep learning has achieved transformative results in image recognition, natural language processing, and other domains, but produces models that are often difficult to interpret. See also: Neural network, Machine learning, Black box.

Deepfake Synthetic media — typically video or audio — in which a person's likeness or voice is convincingly replaced or fabricated using deep learning techniques. Deepfakes pose serious threats in disinformation, non-consensual intimate imagery, fraud, and reputational harm. See also: Synthetic media, Generative AI, Non-consensual intimate imagery.

Demographic Parity A fairness criterion requiring that an AI system's positive decision rate (e.g., loan approval, hiring) be equal across demographic groups. Also called statistical parity. Demographic parity can conflict with other fairness metrics (such as equalized odds) when base rates differ across groups. See also: Equalized odds, Individual fairness, Fairness.

Deontology An ethical theory holding that certain actions are inherently right or wrong regardless of their consequences, based on duties, rights, or rules. Deontological constraints — such as prohibitions on deception or violations of consent — provide important limits on purely consequentialist AI optimization. See also: Categorical imperative, Contractualism, Consequentialism.

Differential Privacy A mathematical framework for sharing information about a dataset while providing strong guarantees that any individual's specific data cannot be inferred. Differential privacy adds carefully calibrated statistical noise to queries or model outputs, providing a rigorous privacy guarantee. See also: Privacy-enhancing technologies, Federated learning, Homomorphic encryption.

Disparate Impact A legal and statistical concept referring to a facially neutral policy or practice that disproportionately disadvantages members of a protected group. Established in Griggs v. Duke Power Co. (1971), the disparate impact doctrine applies to employment and, by extension, to algorithmic systems used in hiring, lending, and housing. See also: Disparate treatment, Demographic parity, Civil rights.

Disparate Treatment Intentional discrimination against individuals because of their membership in a protected class. Unlike disparate impact, disparate treatment requires discriminatory intent. Algorithmic systems can facilitate disparate treatment when protected characteristics are explicitly used as features or when proxy variables are deliberately selected to correlate with protected characteristics. See also: Disparate impact, Civil rights, Proxy variable.

Distribution Shift The phenomenon in which the statistical distribution of data encountered at deployment differs from the distribution of training data, causing model performance to degrade. Distribution shift can occur when the world changes (temporal shift) or when deployment populations differ from training populations. See also: Aggregation bias, Representation bias, Training data.

E

ECOA (Equal Credit Opportunity Act) A U.S. federal law prohibiting discrimination in credit transactions on the basis of race, color, religion, national origin, sex, marital status, age, or receipt of public assistance. The ECOA requires adverse action notices and applies to algorithmic credit scoring and lending systems. See also: Fair Housing Act, Title VII, Adverse action notice.

eGFR (Estimated Glomerular Filtration Rate) A clinical measure of kidney function calculated from blood creatinine levels, age, sex, and historically race. The inclusion of a race-based correction factor in eGFR calculations — which artificially raised scores for Black patients and delayed specialist referrals — became a prominent case study in how embedded assumptions in medical algorithms perpetuate health disparities. See also: Algorithmic bias, Healthcare AI.

Electronic Health Record (EHR) A digital record of a patient's medical history, diagnoses, treatments, medications, and test results. EHR data is widely used to train clinical AI models, but it reflects historical patterns of care, including disparities in treatment and diagnosis across demographic groups. See also: Healthcare AI, Historical bias, Training data.

Equal Opportunity (Fairness) A fairness criterion requiring that among individuals who would qualify for a positive outcome (true positives), the probability of being correctly identified is equal across demographic groups. It focuses on equalizing false negative rates. See also: Equalized odds, Demographic parity, Fairness.

Equalized Odds A fairness criterion proposed by Hardt, Price, and Srebro requiring that a classifier's true positive rate and false positive rate be equal across protected groups. Equalized odds is more demanding than equal opportunity because it constrains both types of error. Formal proofs show that equalized odds, demographic parity, and calibration cannot all be satisfied simultaneously when base rates differ. See also: Demographic parity, Calibration, Chouldechova impossibility.

Ethics Board An organizational body charged with advising on, reviewing, or overseeing the ethical dimensions of AI development and deployment. Ethics boards vary widely in authority, independence, composition, and effectiveness; critics warn that toothless boards can become a form of ethics washing. See also: Governance, Ethics washing, Accountability.

Ethics Washing The practice of using ethical language, commitments, frameworks, or bodies as a public relations strategy — without making substantive changes to harmful practices. Ethics washing substitutes the appearance of ethical responsibility for its substance. See also: Ethics board, Responsible AI, Accountability.

EU AI Act The European Union's comprehensive regulatory framework for artificial intelligence, adopted in 2024, which establishes a risk-based classification of AI systems (unacceptable risk, high risk, limited risk, minimal risk) and imposes varying requirements for transparency, human oversight, data governance, and conformity assessment. The world's first comprehensive binding AI regulation. See also: Brussels Effect, GDPR, High-risk AI.

Explainability The ability of an AI system to provide human-understandable reasons for its outputs or decisions. Explainability is often distinguished from interpretability: interpretability refers to understanding the internal model mechanism, while explainability refers to communicating a decision in a way that is meaningful to stakeholders. See also: Interpretability, XAI, Contrastive explanation, Right to explanation.

Explainable AI (XAI) A field of research and practice focused on developing AI systems whose decisions can be explained to users, regulators, and affected individuals. Key XAI techniques include LIME, SHAP, counterfactual explanations, and decision tree surrogates. See also: LIME, SHAP, Interpretability, Explainability.

F

Facial Recognition A biometric technology that identifies or verifies individuals by analyzing facial features in images or video. Facial recognition systems have demonstrated significant disparate accuracy across demographic groups — performing worst for darker-skinned women — and raise serious concerns about surveillance, civil liberties, and consent. See also: Biometrics, Clearview AI, FRVT, Algorithmic bias.

Fair Housing Act (FHA) A U.S. federal law prohibiting discrimination in the sale, rental, and financing of housing on the basis of race, color, national origin, religion, sex, familial status, or disability. The FHA applies to algorithmic tools used in housing advertising, tenant screening, and mortgage lending. See also: ECOA, Title VII, Algorithmic redlining.

Fairness A multidimensional concept in AI ethics referring to the equitable treatment of individuals and groups by AI systems. There is no single definition: formal fairness metrics (demographic parity, equalized odds, calibration, individual fairness) often conflict, and mathematical proofs establish that they cannot all be satisfied simultaneously. Which definition of fairness to apply is ultimately a political and ethical choice. See also: Demographic parity, Equalized odds, Individual fairness, Counterfactual fairness.

Feature Engineering The process of selecting, transforming, or constructing input variables (features) from raw data to improve model performance. Feature engineering choices embed human judgment about what information is relevant and can inadvertently introduce or perpetuate bias. See also: Proxy variable, Measurement bias, Training data.

Feature Importance A model explanation technique that quantifies the contribution of each input variable to a model's predictions, either globally (across all predictions) or locally (for a specific prediction). Feature importance scores are used in model auditing and explanation but can be misleading when features are correlated. See also: SHAP, LIME, Explainability.

Federated Learning A machine learning approach in which a shared model is trained across multiple decentralized devices or servers holding local data, without transferring raw data to a central server. Federated learning preserves data privacy by keeping individual records local while still enabling model training on distributed data. See also: Differential privacy, Privacy-enhancing technologies.

Feedback Loop A dynamic in which an AI system's outputs influence the data used to train or update subsequent versions of that system, potentially amplifying initial biases. A criminal risk assessment tool that increases incarceration of a demographic group produces more criminal records for that group, reinforcing the model's initial predictions. See also: Historical bias, Algorithmic bias, Distribution shift.

Foundation Model A large AI model trained on broad, diverse data at scale — typically using self-supervised learning — that can be adapted to a wide range of downstream tasks through fine-tuning or prompting. GPT-4, Claude, and Gemini are examples. Foundation models concentrate capabilities (and risks) in a small number of organizations. See also: Large language model, Generative AI, Transfer learning.

FRVT (Face Recognition Vendor Technology) NIST's ongoing evaluation of facial recognition algorithms across a wide range of vendors and conditions. FRVT studies have documented large disparate error rates across demographic groups, providing empirical grounding for civil liberties concerns about facial recognition deployment. See also: Facial recognition, NIST AI RMF, Algorithmic bias.

FTC (Federal Trade Commission) The primary U.S. federal agency with authority over unfair or deceptive trade practices, including many AI-related harms in commercial contexts. The FTC has issued guidance on AI and algorithms, taken enforcement actions related to algorithmic deception, and advocated for algorithmic transparency. See also: UDAP, Consumer protection, Governance.

G

GDPR (General Data Protection Regulation) The European Union's comprehensive data protection law, effective May 2018, establishing broad rights for data subjects and imposing significant obligations on organizations that process personal data, including requirements for lawful basis, purpose limitation, data minimization, and security. Article 22 restricts solely automated decisions with significant effects on individuals. See also: Data subject, Privacy by design, Brussels Effect, Right to explanation.

Generative AI AI systems capable of producing new content — text, images, audio, video, code — that resembles human-created content. Generative AI includes large language models, image generation models (Stable Diffusion, DALL-E), and video generation models. It raises novel concerns around intellectual property, deepfakes, synthetic media, and misinformation. See also: Foundation model, Large language model, Deepfake, Synthetic media.

Ghost Work The largely invisible, low-wage human labor that underlies AI systems — including data annotation, content moderation, and output evaluation. Coined by Mary Gray and Siddharth Suri, the term highlights that AI automation depends on human workers whose contributions are systematically obscured. See also: Data annotation, Labor displacement, Responsible AI.

Global Workspace Theory A scientific theory of consciousness proposing that the brain integrates information from many specialized subsystems through a "global workspace" that broadcasts information widely. Sometimes invoked in debates about AI consciousness to argue that large AI models with broad attention mechanisms may implement functional analogs to conscious processes. See also: Hard problem of consciousness, Integrated Information Theory, AI consciousness.

Governance The systems of rules, practices, norms, incentives, and institutional structures through which authority is exercised and decisions are made. In AI ethics, governance encompasses organizational policies, industry self-regulation, professional standards, and government regulation. See also: AI governance, Accountability, NIST AI RMF.

Gradient Descent The primary optimization algorithm used to train machine learning models. The algorithm iteratively adjusts model parameters in the direction that most reduces a loss function, moving "downhill" on the error surface. The choice of what loss function to minimize is an ethical decision with profound consequences for model behavior. See also: Optimization, Machine learning, Optimization target bias.

H

Hallucination The tendency of large language models to generate confident-sounding but factually incorrect, fabricated, or nonsensical statements. Hallucinations pose serious risks in high-stakes applications (legal, medical, financial) where users may not verify AI outputs. See also: Large language model, Explainability, Accuracy.

Hard Problem of Consciousness Philosopher David Chalmers' formulation of the deep explanatory gap between physical or functional descriptions of brain processes and the subjective experience of "what it is like" to be a conscious creature. The hard problem poses a fundamental challenge to determining whether AI systems could ever be genuinely conscious, as opposed to functionally simulating consciousness. See also: Global Workspace Theory, Integrated Information Theory, Moral patiency.

High-Risk AI (EU AI Act) Under the EU AI Act, a category of AI systems subject to strict obligations including conformity assessments, technical documentation, human oversight, and transparency requirements. High-risk AI includes systems used in biometric identification, critical infrastructure, education, employment, credit, healthcare, law enforcement, border control, and the administration of justice. See also: EU AI Act, Risk assessment, Governance.

Historical Bias Bias in training data that reflects past human discrimination or unequal treatment. When AI models learn from historical data, they tend to replicate and perpetuate the inequities embedded in that history. A hiring algorithm trained on past hiring decisions will learn to prefer candidates resembling those historically hired, who may have been disproportionately white and male. See also: Feedback loop, Algorithmic bias, Training data.

HMDA (Home Mortgage Disclosure Act) A U.S. federal law requiring mortgage lenders to collect and disclose data on loan applications, originations, and denials by race, ethnicity, sex, and geography. HMDA data has been used by researchers and journalists to investigate algorithmic lending discrimination. See also: Fair Housing Act, ECOA, Algorithmic redlining.

Homomorphic Encryption A form of encryption that allows computations to be performed on encrypted data without decrypting it, so that the result, when decrypted, is identical to the result of performing the operation on the plaintext. Homomorphic encryption enables privacy-preserving data analysis and AI computation. See also: Differential privacy, Secure multi-party computation, Privacy-enhancing technologies.

Human-in-the-Loop A system design principle requiring that humans retain the ability to review, override, or meaningfully supervise AI outputs before they become consequential decisions. True human-in-the-loop oversight is undermined by automation bias, time pressure, and high volume — conditions that often prevail in practice. See also: Automation bias, Oversight, Accountability.

I

Impact Assessment A structured process for evaluating the potential effects — harms, benefits, and fairness implications — of a technology, policy, or practice before or during deployment. In AI ethics, impact assessments are increasingly required by regulation (e.g., EU AI Act, some U.S. state laws) and recommended by governance frameworks. See also: Algorithmic impact assessment, NIST AI RMF, Governance.

Individual Fairness A fairness criterion proposed by Dwork et al. requiring that similar individuals (as measured by a task-relevant similarity metric) receive similar predictions or decisions. Individual fairness avoids group-level statistical aggregation but requires defining a morally appropriate similarity metric, which is itself a normative choice. See also: Demographic parity, Counterfactual fairness, Fairness.

Informed Consent The ethical and legal requirement that individuals receive adequate information about how their data will be used and have the opportunity to agree or refuse before data collection or processing occurs. In AI contexts, "consent" obtained through lengthy terms of service contracts that few read is increasingly criticized as inadequate. See also: Privacy by design, Purpose limitation, Data subject.

Integrated Information Theory (IIT) A scientific theory of consciousness proposed by Giulio Tononi, holding that consciousness is identical to integrated information (phi) — the degree to which a system's parts are causally integrated in a way that cannot be decomposed. IIT has been used to argue that some AI architectures have minimal or zero consciousness, while others — in principle — might have more. See also: Global Workspace Theory, Hard problem of consciousness, Moral patiency.

Interpretability The degree to which the internal mechanisms of an AI model can be understood by humans. Interpretability differs from explainability: an interpretable model (like a decision tree) can be directly examined, while an explainability technique for a black-box model provides an approximation. See also: Explainability, Black box, XAI.

Intersectionality A concept introduced by legal scholar Kimberlé Crenshaw to describe the way in which multiple dimensions of identity (race, gender, class, disability, sexuality) interact to create overlapping and compounding forms of discrimination. In AI fairness, intersectionality highlights that models may be fair for each protected group individually while still being discriminatory for people at the intersection of multiple groups. See also: Algorithmic bias, Demographic parity, Civil rights.

J

Justice A foundational concept in ethics and political philosophy concerning the fair distribution of benefits and burdens, equal treatment under the law, and adequate remedies for wrongs. In AI ethics, questions of justice arise in the distribution of AI's benefits and harms across society, and in procedural fairness in automated decision-making. See also: Fairness, Rawls, Capabilities approach.

L

Label Bias Bias introduced when the labels used to train a supervised machine learning model are themselves the product of human judgments that reflect prejudice, inconsistency, or historical discrimination. Criminal recidivism labels based on rearrest (rather than reoffending) introduce label bias because arrest rates are not race-neutral. See also: Data annotation, Historical bias, Training data.

Large Language Model (LLM) A foundation model trained on massive text corpora using self-supervised learning to predict text. LLMs can generate, summarize, translate, and reason about text and, when sufficiently large and fine-tuned, exhibit emergent capabilities. They are the basis of conversational AI systems such as ChatGPT and Claude. See also: Foundation model, Generative AI, Hallucination, Attention mechanism.

LAWS (Lethal Autonomous Weapons Systems) Weapons systems that can select and engage targets without meaningful human intervention, using AI-based perception and decision-making. LAWS raise profound ethical questions about human dignity, accountability for killing, the laws of armed conflict, and the stability of international security. See also: Autonomous weapons, AI safety, Meaningful human control.

LGPD (Lei Geral de Proteção de Dados — Brazil) Brazil's General Data Protection Law, effective 2020, establishing rights for data subjects and obligations for data processors broadly similar to the GDPR. The LGPD reflects the spread of comprehensive data protection regimes beyond the European Union. See also: GDPR, Data subject, Global AI governance.

LIME (Local Interpretable Model-agnostic Explanations) An explainability technique that approximates the behavior of a complex model in the neighborhood of a specific prediction by training a simple interpretable model (e.g., linear regression) on locally perturbed samples. LIME provides locally faithful but potentially globally inconsistent explanations. See also: SHAP, XAI, Explainability.

Logistic Regression A statistical classification algorithm that models the probability of a binary outcome as a function of input features, producing an interpretable model in which coefficients indicate the direction and magnitude of each feature's influence. Logistic regression is widely used in credit scoring and risk assessment and is often valued for its interpretability relative to more complex models. See also: Classification, Interpretability, Machine learning.

M

Machine Learning (ML) A subfield of artificial intelligence in which systems learn to perform tasks from examples (training data) rather than through explicit programming. Machine learning encompasses supervised, unsupervised, and reinforcement learning and underlies most deployed AI applications. See also: Deep learning, Neural network, Training data.

Meaningful Human Control A standard in weapons and autonomous systems ethics requiring that humans retain genuine (not merely nominal) authority over consequential decisions, particularly the use of lethal force. Meaningful human control is undermined when system speed, complexity, or workload prevents genuine deliberation. See also: Human-in-the-loop, LAWS, Oversight.

Measurement Bias Bias arising when the variables used to operationalize constructs of interest are measured in systematically different ways across groups, or when the construct itself is proxied imperfectly. Using credit scores to measure "creditworthiness" introduces measurement bias if credit access has historically been unequal. See also: Proxy variable, Label bias, Historical bias.

Model Card A structured document, proposed by Mitchell et al. (2019), that provides standardized information about a machine learning model including its intended use, training data, performance metrics across demographic groups, and known limitations. Model cards promote transparency and are increasingly expected by responsible AI governance frameworks. See also: Transparency, Algorithmic transparency, Governance.

Model Distillation A technique for training a smaller, more efficient model (the "student") to approximate the behavior of a larger model (the "teacher"). Model distillation enables deployment of capable models on resource-constrained devices but may affect the robustness and fairness of the resulting model. See also: Foundation model, Transfer learning.

Model Extraction An attack in which an adversary queries a model repeatedly to reconstruct a functional approximation of it, potentially stealing intellectual property or enabling subsequent adversarial attacks. See also: Adversarial attack, Cybersecurity.

Moral Luck The philosophical problem that our moral assessments of agents often depend on factors outside their control (e.g., whether an action caused harm through luck). In AI ethics, moral luck arises in questions of accountability: who is responsible when an AI system causes harm through a rare and unforeseeable error? See also: Accountability, Negligence, Strict liability.

Moral Patiency The property of being the kind of entity that can be wronged — that has interests that make moral demands on others. Stones have no moral patiency; humans clearly do; the moral patiency of animals, and potentially AI systems, is contested. See also: Moral status, Hard problem of consciousness, AI consciousness.

Moral Status The standing of an entity as an object of direct moral consideration, such that its interests generate obligations in others. The question of whether advanced AI systems might acquire moral status is a live — if contested — topic in AI ethics and philosophy. See also: Moral patiency, Hard problem of consciousness.

N

Natural Language Processing (NLP) A subfield of AI concerned with enabling computers to understand, generate, and manipulate human language. NLP underlies applications including machine translation, sentiment analysis, chatbots, and large language models. NLP systems can perpetuate linguistic biases present in training corpora. See also: Large language model, Algorithmic bias, Word embeddings.

Negligence A legal tort standard requiring that a defendant failed to exercise the care that a reasonable person would exercise under the circumstances, and that this failure caused harm. In AI liability, negligence claims require establishing the standard of care for AI development and deployment — a developing area of law. See also: Strict liability, Tort, Accountability.

Neural Network A computational model loosely inspired by the brain's structure, consisting of interconnected layers of nodes (neurons) that process information. Neural networks learn by adjusting the weights of connections between nodes during training. Deep neural networks with many layers are the foundation of modern AI applications. See also: Deep learning, Machine learning, Attention mechanism.

NIST AI RMF (AI Risk Management Framework) A voluntary framework published by the U.S. National Institute of Standards and Technology in 2023 providing guidance for organizations to manage AI risks across the AI lifecycle. The RMF organizes risk management around four functions: Govern, Map, Measure, and Manage. See also: Governance, Impact assessment, Responsible AI.

Non-Consensual Intimate Imagery (NCII) Sexually explicit images or video depicting a person distributed without their consent, often produced using deepfake or image synthesis technology. NCII causes severe psychological harm and is increasingly recognized as a form of sexual abuse and a target of criminal legislation. See also: Deepfake, Synthetic media, Generative AI.

O

Optimization The mathematical process of finding parameter values that minimize (or maximize) an objective function. Machine learning training is a form of optimization. The choice of what to optimize is an ethical decision: optimizing for engagement maximization, for example, can amplify outrage and disinformation. See also: Gradient descent, Optimization target bias.

Optimization Target Bias Bias arising when the metric or objective function a model is trained to optimize is a poor or biased proxy for the socially desirable outcome. A hiring model optimized to predict "performance" measured by manager ratings will embed whatever biases exist in the manager rating process. See also: Optimization, Proxy variable, Measurement bias.

ORCA (O'Neil Risk Consulting) The consulting firm founded by mathematician and author Cathy O'Neil (author of Weapons of Math Destruction) that audits algorithms for fairness and bias. ORCA represents a model of third-party algorithmic auditing as a profession. See also: Auditing AI systems, Algorithmic accountability.

Oversight The processes by which humans monitor, review, and retain control over AI systems and their consequences. Effective oversight requires access to information about system performance, institutional authority to intervene, and protection from conflicts of interest. See also: Human-in-the-loop, Accountability, Governance.

P

Participatory Design An approach to technology design that actively involves end users and affected communities in the design process, rather than treating them as passive recipients of decisions made by technical experts. Participatory design is increasingly advocated as an approach to building fairer, more legitimate AI systems. See also: Stakeholder, Impact assessment, Care ethics.

Principal-Agent Problem An economic and organizational theory concept describing conflicts of interest that arise when one party (the agent) acts on behalf of another (the principal) and the agent's incentives diverge from the principal's interests. In AI governance, the principal-agent problem arises in relationships between organizations and their AI vendors, or between AI developers and the public. See also: Governance, Accountability, Ethics board.

Privacy by Design A framework for embedding privacy protections into the design of systems and processes from the outset, rather than as an afterthought. Privacy by design principles include data minimization, purpose limitation, security by default, and user control. See also: Data minimization, Purpose limitation, GDPR.

Privacy-Enhancing Technologies (PETs) A family of technical tools designed to protect personal privacy while enabling useful computation, including differential privacy, federated learning, homomorphic encryption, and secure multi-party computation. PETs enable privacy-preserving AI development and analysis. See also: Differential privacy, Federated learning, Homomorphic encryption.

Proxy Variable A variable used as a substitute for a characteristic that is difficult or impossible to measure directly, or that is legally impermissible to use. In algorithmic systems, apparently neutral variables (e.g., zip code, school attended, name) can serve as proxies for race or other protected characteristics, enabling discrimination without explicit use of protected attributes. See also: Algorithmic redlining, Disparate impact, Feature engineering.

Psychological Safety A workplace condition in which team members feel safe to speak up, raise concerns, and challenge assumptions without fear of punishment. Psychological safety is essential for effective AI ethics oversight within organizations, enabling engineers, product managers, and others to flag ethical concerns. See also: Whistleblower, Ethics board, Governance.

Purpose Limitation The privacy principle that data collected for one specific, stated purpose should not be used for a different incompatible purpose. Purpose limitation constrains the repurposing of personal data for AI training or for uses not anticipated at the time of collection. See also: Data minimization, Privacy by design, GDPR.

R

Recidivism The tendency of a convicted criminal to re-offend after release. Recidivism prediction is a central application of algorithmic risk assessment in criminal justice, and a focal point for AI bias research — particularly following the ProPublica analysis of COMPAS. See also: COMPAS, Risk assessment (criminal justice), Algorithmic bias.

Red-Teaming A structured adversarial testing process in which a designated team attempts to find failures, vulnerabilities, or harmful outputs in an AI system before deployment. Red-teaming is used in both cybersecurity and AI safety contexts. See also: Adversarial attack, AI safety, Auditing AI systems.

Reinforcement Learning from Human Feedback (RLHF) A training technique in which human evaluators rank or rate model outputs, and those preferences are used to train a reward model, which then guides further model training via reinforcement learning. RLHF has been central to aligning large language models with human preferences. See also: Large language model, Foundation model, AI safety.

Representation Bias Bias arising when training data fails to adequately represent the full range of the population or phenomenon of interest. Models trained on data that overrepresents certain groups while underrepresenting others will perform poorly for underrepresented groups. See also: Aggregation bias, Historical bias, Training data.

Responsible AI A broad framework and organizational commitment to developing and deploying AI in ways that are fair, transparent, accountable, safe, and beneficial. Responsible AI integrates ethical values into technical and business practices, and is distinguished from ethics washing by substantive implementation. See also: AI ethics, Governance, Ethics washing.

Risk Assessment (Criminal Justice) The use of structured instruments or algorithmic tools to estimate the probability that an individual will commit future offenses, used to inform decisions about bail, sentencing, parole, and supervision. Algorithmic risk assessments have been criticized for racial bias, opacity, and substitution of actuarial prediction for individualized justice. See also: COMPAS, Recidivism, Disparate impact.

Rule of Law The principle that governmental authority is exercised in accordance with established, publicly known, and consistently applied legal rules, rather than arbitrarily. Automated governmental decision-making raises rule-of-law concerns when it operates opaquely, without adequate procedural safeguards, or in ways that cannot be meaningfully challenged. See also: Due process, Accountability, Governance.

S

SaMD (Software as a Medical Device) Software that is intended to be used for medical purposes — diagnosis, prevention, monitoring, treatment — and is therefore subject to regulatory oversight as a medical device. In the U.S., SaMD is regulated by the FDA; the regulatory framework for AI-based SaMD is actively evolving. See also: Healthcare AI, High-risk AI, FDA.

Secure Multi-Party Computation (SMPC) A cryptographic method that allows multiple parties to jointly compute a function over their combined data without any party revealing its private inputs to the others. SMPC enables collaborative AI training on sensitive data without centralizing it. See also: Homomorphic encryption, Federated learning, Privacy-enhancing technologies.

Sensitive Category Data Personal data that warrants heightened protection because of its particular sensitivity or the severity of harm its misuse can cause — including racial or ethnic origin, political opinions, religious beliefs, health data, sexual orientation, genetic data, and biometric data. Most comprehensive privacy laws impose stricter controls on sensitive category data. See also: GDPR, Privacy by design, Biometrics.

SHAP (SHapley Additive exPlanations) An explainability technique grounded in cooperative game theory (Shapley values) that assigns each feature a contribution score for each prediction, representing the feature's average marginal contribution across all possible orderings of features. SHAP provides a theoretically principled and globally consistent approach to feature importance. See also: LIME, Feature importance, XAI.

Social Credit System A set of Chinese government initiatives using AI and big data to score individual and corporate behavior across financial, legal, and social dimensions. The social credit system has been widely criticized as a form of mass surveillance and behavioral control. It is often cited as a cautionary example in AI ethics debates about surveillance and social sorting. See also: Surveillance capitalism, Facial recognition, AI governance.

Stakeholder Any individual, group, or organization that is affected by or can affect an AI system — including those who develop, deploy, use, and are subject to the system. Comprehensive AI governance requires identifying and engaging all relevant stakeholders, including those who are disadvantaged and whose voices may not be heard through market mechanisms. See also: Participatory design, Impact assessment, Governance.

Strict Liability A legal doctrine imposing liability for harm regardless of fault or intent — the defendant is responsible simply because their activity caused the harm. Strict liability is appropriate for ultrahazardous activities and may be advocated for certain high-risk AI applications, removing the need to prove negligence. See also: Negligence, Tort, Accountability.

Surveillance Capitalism A term coined by Shoshana Zuboff for an economic system in which human experience is treated as a raw material for behavioral prediction and modification, primarily for commercial ends. Surveillance capitalism is the business model of many large technology platforms. See also: Cambridge Analytica, Privacy by design, Data colonialism.

Synthetic Data Data that is artificially generated by algorithms (often AI models) rather than collected from real-world events. Synthetic data can augment training datasets, preserve privacy by avoiding the use of real personal data, and address representation bias — but can also amplify existing model biases if generated from biased models. See also: Data augmentation, Differential privacy, Generative AI.

Synthetic Media Any media content — text, images, audio, video — generated or substantially altered by AI. Synthetic media encompasses both beneficial applications (creative tools, content accessibility) and harmful ones (deepfakes, disinformation). See also: Deepfake, Generative AI, Non-consensual intimate imagery.

T

Title VII (of the Civil Rights Act of 1964) The principal U.S. federal law prohibiting employment discrimination on the basis of race, color, religion, sex, or national origin. Title VII encompasses both disparate treatment and disparate impact theories of discrimination and applies to algorithmic hiring, screening, and evaluation systems. See also: ECOA, Fair Housing Act, Disparate impact.

Tort A civil wrong (other than breach of contract) for which a court may award damages. Tort law is a primary vehicle through which AI-related harms can be remedied, via theories including negligence, strict liability, and products liability. See also: Negligence, Strict liability, Accountability.

Training Data The dataset used to teach a machine learning model to perform a task. The quality, representativeness, and ethical provenance of training data profoundly influence model performance, fairness, and safety. See also: Historical bias, Representation bias, Label bias, Data annotation.

Transfer Learning A machine learning technique in which a model trained on one task or dataset is adapted (fine-tuned) to a different but related task. Transfer learning enables powerful capabilities to be built with less task-specific data, but can also transfer biases embedded in the original training. See also: Foundation model, Model distillation, Deep learning.

Transparency The quality of making information about an AI system — its design, training data, decision criteria, performance, and governance — accessible and understandable to relevant parties. Transparency enables accountability but must be balanced against legitimate confidentiality interests (trade secrets, security). See also: Algorithmic transparency, Explainability, Governance.

True Positive / False Positive / True Negative / False Negative The four possible outcomes of a binary classification: a true positive (TP) is a correct positive prediction; a false positive (FP) is an incorrect positive prediction; a true negative (TN) is a correct negative prediction; a false negative (FN) is an incorrect negative prediction. These metrics are foundational to AI fairness analysis and are combined differently in different fairness criteria. See also: Accuracy, Calibration, Equalized odds.

U

Ubuntu An African philosophical concept expressing the idea that personhood is fundamentally relational — "I am because we are." Ubuntu ethics prioritizes community, interdependence, and collective flourishing over individual utility. Applied to AI ethics, ubuntu challenges methodological individualism and highlights the importance of community impact and communal values in AI governance. See also: Confucian ethics, Care ethics, Capabilities approach.

UDAP (Unfair and Deceptive Acts and Practices) Legal prohibitions, enforced primarily by the FTC under Section 5 of the FTC Act and analogous state laws, against business practices that are unfair or deceptive to consumers. UDAP authority has been applied to algorithmic deception, biased AI systems, and unfair data practices. See also: FTC, Consumer protection, Governance.

Universal Basic Income (UBI) A policy proposal for providing all citizens with a regular unconditional cash payment, often discussed as a response to potential large-scale job displacement from automation and AI. Debates about UBI reflect deeper questions about the appropriate distribution of productivity gains from AI. See also: Labor displacement, AI and employment, Justice.

V

Veil of Ignorance A thought experiment proposed by John Rawls in which individuals choose the principles governing their society without knowing their own position in it — their race, wealth, talents, or other circumstances. Rawls argued that rational agents behind the veil of ignorance would choose principles protecting the worst-off. Applied to AI ethics, the veil of ignorance asks what AI governance rules we would choose if we did not know whether we would be among those benefiting from or harmed by AI systems. See also: Contractualism, Rawls, Justice.

Virtue Ethics An ethical tradition — tracing to Aristotle — that focuses on the character of moral agents rather than on rules or consequences. Virtue ethics asks what a person of good character would do, and emphasizes developing virtues such as practical wisdom (phronesis), honesty, courage, and justice. Applied to AI, virtue ethics asks what character traits organizations and developers should cultivate. See also: Care ethics, Deontology, Consequentialism.

W

Whistleblower An individual who reports illegal, unethical, or dangerous conduct by their employer or another organization, often at personal risk. In AI ethics, whistleblowers have been important in revealing harmful algorithmic practices. Effective AI governance requires robust legal protections for AI ethics whistleblowers. See also: Psychological safety, Accountability, Governance.

Word Embeddings Vector representations of words in a high-dimensional space, learned from patterns of co-occurrence in large text corpora. Word embeddings capture semantic and syntactic relationships but also encode cultural biases: Caliskan et al. (2017) demonstrated that standard word embeddings exhibit human-like racial and gender biases. See also: Natural language processing, Large language model, Algorithmic bias.

X

XAI (Explainable AI) See: Explainable AI.

Z

Zero-Shot Learning A machine learning setting in which a model is asked to perform tasks or classify items it has not encountered in training, relying on learned relationships or structured descriptions of new categories. Zero-shot capability is a key feature of large language models and foundation models, enabling generalization without task-specific training data. See also: Foundation model, Large language model, Transfer learning.

This glossary covers the principal terms used throughout the textbook. For further detail on any concept, readers are referred to the chapter in which the term is most extensively treated, and to the bibliography for primary sources. Where mathematical definitions of fairness criteria are required, readers are directed to Barocas, Hardt, and Narayanan, Fairness and Machine Learning (2023).