Appendix F: Glossary

All key terms from the textbook, organized alphabetically. Each entry includes the chapter and section where the term is first introduced, a plain-language definition, and cross-references to related terms where helpful. Terms marked with a dagger (†) are threshold concepts — ideas that fundamentally change how you understand the subject once you grasp them.

A

Adaptive learning (Ch. 16, §16.3): An educational technology approach that adjusts content, pace, or difficulty based on a student's performance in real time. Uses AI to personalize the learning experience. See also: intelligent tutoring system, personalized learning.

Adversarial example (Ch. 6, §6.6): An input deliberately modified — often in ways imperceptible to humans — to cause an AI system to make an incorrect classification. A classic example: adding carefully calculated pixel-level noise to an image of a panda causes a classifier to identify it as a gibbon with high confidence. See also: adversarial input, edge case.

Adversarial input (Ch. 8, §8.1): Any input designed to exploit weaknesses in an AI system, causing it to fail or produce incorrect outputs. See also: adversarial example, red-teaming.

Aggregation bias (Ch. 9, §9.2): Bias that occurs when a single model is used for groups with different underlying characteristics, and the model performs well on average but poorly for specific subgroups. See also: representation bias, fairness-accuracy trade-off.

AI agents (Ch. 21, §21.1): AI systems capable of taking autonomous actions in the world — browsing the web, executing code, interacting with other software, and making multi-step decisions to accomplish goals. Distinguished from passive AI by their ability to act, not just respond.

AI effect (Ch. 1, §1.4): The phenomenon in which, once a machine can perform a task previously considered a hallmark of intelligence, people redefine intelligence to exclude that task. As computer scientist Larry Tesler quipped: "AI is whatever hasn't been done yet." See also: narrow AI, general AI.

AI geopolitics (Ch. 19, §19.1): The competition between nations and blocs — particularly the United States, China, and the European Union — to lead in AI development, set global standards, and leverage AI for economic and strategic advantage.

AI literacy (Ch. 1, §1.6): The ability to critically evaluate, effectively use, and meaningfully participate in decisions about AI systems. Encompasses technical understanding, ethical reasoning, and civic engagement. This textbook treats AI literacy as a civic skill essential for democratic participation.

AI literacy framework (Ch. 21, §21.4): A structured approach for evaluating any AI system, incorporating technical analysis, ethical assessment, stakeholder impact mapping, and governance review. The FACTS Framework introduced in Chapter 1 is one example.

AI safety (Ch. 20, §20.2): The field of research dedicated to ensuring AI systems behave reliably, robustly, and in accordance with human intentions. Encompasses both near-term concerns (robustness, reliability, misuse prevention) and long-term concerns (existential risk).

AI winter (Ch. 2, §2.2): A period of reduced funding, interest, and progress in AI research, typically following a cycle of overhyped promises and disappointing results. Two major AI winters occurred: roughly 1974–1980 and 1987–1993. See also: hype cycle.

Algorithm (Ch. 1, §1.2): A step-by-step procedure for solving a problem or performing a computation. In the AI context, algorithms are the mathematical procedures that models use to learn from data and make predictions.

Algorithmic accountability (Ch. 17, §17.4): The principle that organizations deploying AI systems should be answerable for the outcomes those systems produce, including unintended harms. See also: accountability gap, algorithmic impact assessment.

Algorithmic assessment (Ch. 16, §16.4): The use of AI systems to evaluate student work, including automated essay scoring, plagiarism detection, and competency assessment. See also: automated essay scoring, AI proctoring.

Algorithmic bias (Ch. 9, §9.1): Systematic and repeatable errors in an AI system that create unfair outcomes, such as favoring one group over another. May arise from biased training data, flawed model design, or problematic deployment context. See also: historical bias, representation bias, measurement bias.

Algorithmic impact assessment (Ch. 13, §13.6; Ch. 17, §17.5): A structured evaluation conducted before deploying an AI system in a high-stakes context, assessing potential harms, affected populations, fairness, and accountability mechanisms.

Algorithmic management (Ch. 10, §10.5): The use of algorithms and AI to direct, evaluate, and discipline workers — particularly prevalent in gig economy platforms. Includes automated scheduling, performance scoring, and work assignment.

Algorithmic monoculture (Ch. 19, §19.5): The risk that a small number of AI systems, developed in a small number of countries, become globally dominant, homogenizing perspectives and reducing diversity in how information is processed and decisions are made.

† Alignment problem (Ch. 20, §20.1): The challenge of ensuring AI systems do what humans actually want — not just what they are technically instructed to optimize for. Specifying human values and intentions in formal terms that a machine can follow is far harder than it appears. See also: specification gaming, reward hacking, value alignment.

Annotation (Ch. 4, §4.3): The process of adding labels or metadata to data, typically performed by human workers. In AI, annotations provide the ground truth that supervised learning models learn from. See also: data labeling, ground truth.

Artificial general intelligence (AGI) (Ch. 1, §1.3; Ch. 21, §21.1): A hypothetical AI system with the flexibility, adaptability, and breadth of human cognition — able to learn new tasks, transfer knowledge across domains, and reason about novel situations. As of 2026, AGI does not exist. See also: narrow AI.

Artificial intelligence (Ch. 1, §1.2): The design of computer systems that perform tasks typically associated with human intelligence — such as recognizing patterns, making predictions, understanding language, or making decisions. An umbrella term for dozens of different techniques.

Automated essay scoring (Ch. 16, §16.4): AI systems that evaluate and score written work, typically using natural language processing to assess features like structure, vocabulary, grammar, and argumentation.

Automation (Ch. 10, §10.1): The use of technology to perform tasks previously done by humans. In the AI context, automation refers specifically to AI systems taking over specific tasks within a job. See also: augmentation.

Automation bias (Ch. 8, §8.4): The tendency for humans to over-trust the outputs of automated systems, even when those outputs are incorrect. A well-documented cognitive bias that can lead people to accept AI recommendations without sufficient scrutiny. See also: confidence score, calibration.

Augmentation (Ch. 10, §10.1): The use of AI to enhance human capabilities rather than replace human workers. In an augmentation model, AI handles routine tasks while humans focus on judgment, creativity, and relationship-building. See also: automation, human-AI collaboration.

B

Backpropagation (Ch. 2, §2.3; Ch. 3, §3.6): The algorithm used to train neural networks by calculating how much each parameter contributed to the overall error and adjusting the parameters accordingly. Popularized in a 1986 paper by Rumelhart, Hinton, and Williams.

Benchmark (Ch. 2, §2.4): A standardized dataset or task used to evaluate and compare AI models. Examples include ImageNet (computer vision), GLUE (language understanding), and MMLU (general knowledge). See Appendix A for limitations of benchmark-based evaluation.

Bias audit (Ch. 9, §9.5): A systematic evaluation of an AI system for evidence of bias across demographic groups. May be conducted internally by the system's developers, by external auditors, or by independent researchers. See also: algorithmic impact assessment.

Biometric data (Ch. 12, §12.3): Data derived from physical or behavioral characteristics of an individual, including fingerprints, facial geometry, iris patterns, voice, and gait. Biometric data is increasingly used for identification and surveillance.

Brussels effect (Ch. 19, §19.3): The phenomenon in which EU regulations become de facto global standards because companies find it simpler to comply globally than to maintain separate products for different jurisdictions. Named after the EU's regulatory influence from Brussels.

C

Calibration (Ch. 8, §8.4; Ch. 9, §9.3): In AI, the degree to which a system's confidence scores match actual probabilities. A well-calibrated system that reports 80% confidence should be correct approximately 80% of the time. In the fairness context, calibration means a given risk score corresponds to the same probability of the outcome regardless of group membership. See also: confidence score.

Carbon footprint (of AI) (Ch. 18, §18.1): The total greenhouse gas emissions associated with training and running AI models, including the electricity consumed by computation and the embodied carbon in hardware manufacturing. See also: green AI, training vs. inference energy.

Cascading failure (Ch. 8, §8.5): A failure mode in which an error in one AI system propagates through interconnected systems, producing compounding harms. See also: graceful degradation.

Chain-of-thought prompting (Ch. 14, §14.2): A prompting technique in which the user instructs the language model to show its reasoning step by step before arriving at an answer. Often improves accuracy on complex reasoning tasks. See also: zero-shot prompting, few-shot prompting.

Chilling effect (Ch. 12, §12.4): The phenomenon in which people alter their behavior — particularly their speech and political activity — because they believe they are being watched or monitored, even if no specific consequences have occurred. See also: panopticon effect, surveillance capitalism.

Clinical decision support (Ch. 15, §15.1): AI systems designed to assist healthcare providers in making clinical decisions, such as diagnosing conditions, recommending treatments, or flagging potential drug interactions.

Collaborative filtering (Ch. 7, §7.2): A recommendation technique that suggests items based on what similar users liked. If users A and B have similar taste and user A liked item X, the system recommends item X to user B.

Compute divide (Ch. 19, §19.4): The gap between organizations and countries that can afford the massive computing infrastructure needed to develop frontier AI models and those that cannot.

Confidence score (Ch. 8, §8.4): A numerical value produced by an AI system indicating how certain it is about a prediction or classification. A confidence score of 0.95 does not necessarily mean the prediction is correct — it means the model assigns high probability to that output. See also: calibration, automation bias.

Conformity assessment (Ch. 13, §13.6): Under the EU AI Act, a process by which AI systems in the high-risk category must demonstrate compliance with requirements for safety, transparency, and human oversight before being placed on the market.

Consent fatigue (Ch. 12, §12.2): The phenomenon in which individuals stop reading or meaningfully engaging with consent forms and privacy policies because they encounter so many of them.

Constitutional AI (Ch. 20, §20.4): An approach to AI alignment in which a model is trained to follow a set of principles (a "constitution") that guide its behavior. Developed by Anthropic as an alternative to pure RLHF.

Content-based filtering (Ch. 7, §7.2): A recommendation technique that suggests items similar in characteristics to items a user has previously liked, based on the features of the items themselves rather than the behavior of other users.

Convolutional neural network (CNN) (Ch. 6, §6.2): A type of neural network designed for processing grid-structured data like images. CNNs use filters that slide across the image to detect features at increasingly abstract levels — edges, shapes, objects. See also: neural network, feature.

D

Data broker (Ch. 12, §12.2): A company that collects, aggregates, and sells personal data about individuals, often without those individuals' knowledge or meaningful consent.

Data colonialism (Ch. 19, §19.5): The extraction of data from communities — particularly in the Global South — by powerful technology companies in ways that replicate colonial patterns of resource extraction, with benefits flowing primarily to the extractors.

† Data is never neutral (Ch. 4, §4.4): The threshold concept that all datasets reflect the choices, assumptions, biases, and power structures of the people and institutions that created them. Data does not passively record reality — it encodes a particular version of it. See also: ghost data, historical bias.

Data labeling (Ch. 4, §4.3): The process of assigning categories, tags, or annotations to data points so that a supervised learning model can learn the relationship between inputs and desired outputs. Often performed by low-paid workers in developing countries. See also: annotation, ground truth.

Data localization (Ch. 19, §19.5): Laws or policies requiring that data generated within a country be stored and processed within that country's borders. Motivated by concerns about sovereignty, privacy, and security.

Data minimization (Ch. 12, §12.5): The principle of collecting only the minimum amount of personal data necessary for a specified purpose — a core requirement of GDPR and other privacy frameworks.

Data provenance (Ch. 4, §4.5): The documented history of a dataset — where it came from, who collected it, how it was processed, and what transformations it has undergone. See also: datasheets for datasets.

Deep learning (Ch. 2, §2.4; Ch. 3, §3.6): A subset of machine learning that uses neural networks with many layers (hence "deep") to learn hierarchical representations of data. The approach that powers most modern AI systems, including language models, computer vision, and speech recognition.

Deepfake (Ch. 6, §6.5): Synthetic media (typically video or audio) generated using AI to make it appear that a person said or did something they did not. Raises concerns about misinformation, identity theft, and trust in media.

Demographic parity (Ch. 9, §9.3): A fairness criterion requiring that an AI system make positive classifications at the same rate across demographic groups (e.g., the same proportion of male and female applicants are selected). Also called "statistical parity." See also: equalized odds, calibration.

Deskilling (Ch. 16, §16.6): The process by which workers lose skills they no longer practice because AI or automation handles those tasks. In education, the concern that AI tools will prevent students from developing foundational skills.

Diagnostic AI (Ch. 15, §15.1): AI systems designed to assist in medical diagnosis, typically by analyzing medical images (X-rays, CT scans, pathology slides) or patient data to identify conditions or diseases.

Diffusion model (Ch. 11, §11.2): A type of generative AI model that creates images by learning to reverse a gradual noising process. Starting from random noise, the model iteratively removes noise to produce a coherent image. The architecture behind Stable Diffusion, DALL-E, and Midjourney.

Digital divide (educational) (Ch. 16, §16.5): The gap between students who have access to AI tools, reliable internet, and modern devices and those who do not — a gap that mirrors and can amplify existing socioeconomic inequalities.

Digital footprint (Ch. 12, §12.2): The trail of data a person creates through their online activity — searches, purchases, social media posts, location data, browsing history, and more.

Digital sovereignty (Ch. 19, §19.5): A nation or region's capacity to control the digital infrastructure, data, and AI systems operating within its borders, rather than depending on foreign technology companies.

Disparate impact (Ch. 9, §9.4; Ch. 17, §17.3): A legal concept referring to practices that are neutral on their face but disproportionately harm a protected group. In AI, a system can have disparate impact even without any explicit discriminatory intent.

Distributional shift (Ch. 8, §8.3): A change between the data the model was trained on and the data it encounters in deployment, causing performance to degrade. Also called "dataset shift" or "covariate shift." See also: generalization.

Due process (Ch. 17, §17.3): A constitutional principle (in the U.S.) requiring fair procedures before the government can deprive someone of life, liberty, or property. Procedural due process concerns the fairness of the process; substantive due process concerns the fundamental rights at stake.

Durable frameworks (Ch. 21, §21.4): Analytical approaches and conceptual tools for evaluating AI that remain useful even as specific technologies change. A core theme of this textbook: technologies evolve rapidly, but frameworks for thinking about them endure.

E

Edge case (Ch. 8, §8.1): An unusual or extreme input that falls at the boundaries of what a model was trained to handle. AI systems often fail on edge cases because they are underrepresented in training data.

Embodied carbon (Ch. 18, §18.2): The greenhouse gas emissions associated with manufacturing hardware — mining raw materials, fabricating chips, assembling servers — as distinct from the emissions from operating the hardware.

Equal protection (Ch. 17, §17.3): A constitutional principle (in the U.S., under the 14th Amendment) prohibiting the government from denying any person equal protection of the laws. AI systems used in government contexts raise equal protection concerns when they treat demographic groups differently.

Equalized odds (Ch. 9, §9.3): A fairness criterion requiring that an AI system's true positive rate and false positive rate be equal across demographic groups. In other words, the system should be equally accurate (and equally inaccurate) for everyone. See also: demographic parity, calibration.

E-waste (Ch. 18, §18.2): Discarded electronic equipment, including the specialized hardware (GPUs, TPUs, servers) used to train and run AI models. AI's demand for the latest hardware contributes to growing e-waste.

Existential risk (x-risk) (Ch. 20, §20.3): The risk that advanced AI could pose threats to the continued existence of humanity or cause irreversible global-scale catastrophe. A subject of intense debate within the AI community.

Expert system (Ch. 2, §2.3): A type of AI system, popular in the 1980s, that encodes expert knowledge as a set of if-then rules. Expert systems were transparent and interpretable but brittle and expensive to maintain.

F

Face embedding (Ch. 6, §6.4): A numerical representation of a face's geometric features — distances between eyes, nose shape, jawline — computed by a neural network. Used in facial recognition to match faces against a database.

Facial recognition (Ch. 6, §6.4; Ch. 12, §12.3): AI technology that identifies or verifies individuals based on their facial features. One of the most debated AI technologies due to its surveillance capabilities and documented performance disparities across skin tones and genders.

† Fairness is not a single metric (Ch. 9, §9.3): The threshold concept that different mathematical definitions of fairness — demographic parity, equalized odds, calibration — are mutually incompatible under most real-world conditions. Choosing which definition to optimize is a value judgment, not a technical decision.

Fairness-accuracy trade-off (Ch. 9, §9.5): The observation that improving an AI system's fairness on one metric often comes at a cost to overall accuracy, or to fairness on a different metric.

Fairness through unawareness (Ch. 9, §9.2): The (flawed) approach of removing protected characteristics (race, gender, etc.) from a model's inputs in the hope of preventing discrimination. Fails because proxy variables can encode the same information indirectly. See also: proxy variable.

FDA clearance vs. FDA approval (Ch. 15, §15.5): Two distinct regulatory pathways for medical devices (including AI-based devices) in the United States. Clearance (510(k)) requires showing substantial equivalence to an existing device; approval (PMA) requires more rigorous clinical evidence.

Feature (Ch. 3, §3.1): A measurable property of the data that a machine learning model uses to make predictions. In a spam filter, features might include the presence of certain words, the sender's address, or the time of day. In image recognition, features are visual patterns like edges, textures, and shapes.

Feedback loop (Ch. 7, §7.6): A process in which an AI system's outputs influence the data that will be used to retrain or evaluate the system, potentially amplifying biases or errors over time. See also: runaway feedback loop.

Few-shot prompting (Ch. 14, §14.2): A prompting technique in which the user provides a small number of examples of the desired input-output pattern before asking the model to perform the task. See also: zero-shot prompting, chain-of-thought prompting.

Fine-tuning (Ch. 5, §5.4): The process of further training a pre-trained model on a specific dataset or task to improve its performance in a particular domain. Like taking a generally educated person and giving them specialized professional training.

G

Generalization (Ch. 3, §3.4): A model's ability to perform well on new, unseen data — not just the data it was trained on. The central goal of machine learning. See also: overfitting, underfitting.

General AI (strong AI / AGI) (Ch. 1, §1.3): See artificial general intelligence.

General-purpose AI model (Ch. 13, §13.2): Under the EU AI Act, an AI model trained on broad data that can perform a wide range of tasks. Subject to specific transparency and safety requirements.

Generative Adversarial Network (GAN) (Ch. 11, §11.2): A model architecture consisting of two neural networks — a generator that creates synthetic content and a discriminator that tries to distinguish synthetic from real. Through competition, the generator learns to produce increasingly realistic outputs.

Generative AI (Ch. 11, §11.1): AI systems that create new content — text, images, music, video, code — rather than analyzing or classifying existing content. Includes large language models, diffusion models, and GANs.

Ghost data (Ch. 4, §4.4): Information that is absent from a dataset because the people, events, or conditions it would describe were never measured, recorded, or included. Ghost data is invisible by definition but can significantly affect AI performance.

Graceful degradation (Ch. 8, §8.6): The ability of a system to maintain partial functionality when some components fail or when inputs fall outside expected parameters, rather than failing completely and unpredictably.

Green AI (Ch. 18, §18.5): A movement advocating for more energy-efficient AI research and development — prioritizing model efficiency, smaller models, and reduced environmental impact. See also: model efficiency, rebound effect.

Ground truth (Ch. 4, §4.3): The correct answer or accurate real-world information against which an AI system's predictions are measured. In supervised learning, ground truth is typically provided by human labels.

H

Hallucination (Ch. 8, §8.2): In the AI context, the generation of plausible-sounding but factually incorrect or entirely fabricated content by a language model. Occurs because LLMs predict the most probable next tokens rather than verifying factual accuracy. See also: confidence score.

Hard law (Ch. 13, §13.5): Legally binding regulations and statutes that are enforceable by courts and government agencies. Contrasted with soft law.

Health equity (Ch. 15, §15.3): The principle that everyone should have a fair and just opportunity to be as healthy as possible, which requires addressing social, economic, and environmental disadvantages. AI systems in healthcare can either advance or undermine health equity.

Heuristic (Ch. 1, §1.2): A practical rule of thumb or mental shortcut used to make decisions or solve problems quickly, without guaranteed optimality. AI systems often use heuristics when exact solutions are computationally infeasible.

High-risk AI system (Ch. 13, §13.2): Under the EU AI Act, an AI system used in a context where failure could harm health, safety, or fundamental rights — including systems used in education, employment, law enforcement, and critical infrastructure. Subject to strict regulatory requirements.

Historical bias (Ch. 9, §9.2): Bias in AI systems that arises from historical patterns of inequality encoded in training data. Even if the data accurately reflects historical reality, a model trained on that data will reproduce historical inequities. See also: algorithmic bias.

Human-AI collaboration (Ch. 14, §14.6): An approach to using AI in which humans and AI systems work together, with each contributing their respective strengths — AI for pattern recognition and data processing, humans for judgment, context, and ethical reasoning.

Hype cycle (Ch. 2, §2.6): A recurring pattern in AI history: new capabilities generate unrealistic expectations, which lead to disappointment when those expectations are not met, followed by a quieter period of genuine progress.

I

Inference (Ch. 12, §12.2): In the privacy context, the process by which AI systems deduce sensitive information about individuals from seemingly innocuous data. For example, inferring health conditions from purchasing patterns or political views from music preferences.

Innovation principle (Ch. 13, §13.5): The regulatory philosophy that laws should not impede technological innovation and that potential benefits should be weighed against potential harms before regulation is imposed. Often contrasted with the precautionary principle.

Intelligent tutoring system (ITS) (Ch. 16, §16.1): A computer system that provides personalized instruction and feedback to learners, adapting to individual student needs. A concept that predates modern AI but has been revitalized by language models.

Interpretability (Ch. 7, §7.5): The degree to which a human can understand why an AI system made a particular decision. Simpler models (like decision trees) are more interpretable; complex models (like deep neural networks) are less so. See also: explainability in medicine.

K

Knowledge engineering (Ch. 2, §2.3): The process of eliciting, organizing, and encoding expert knowledge into an AI system — the primary task involved in building expert systems. Labor-intensive and brittle.

L

Label (Ch. 3, §3.1): In supervised learning, the known correct answer associated with a training example. A dataset of emails labeled "spam" or "not spam" provides labels for a spam classifier. See also: feature, annotation.

Large language model (LLM) (Ch. 5, §5.1): A neural network trained on massive amounts of text data to predict the next token in a sequence. LLMs like GPT, Claude, Gemini, and Llama can generate coherent text, answer questions, write code, and perform many language tasks. See also: transformer, next-token prediction.

Learning analytics (Ch. 16, §16.4): The collection, analysis, and use of data about learners and their contexts for the purpose of understanding and optimizing learning and the environments in which it occurs.

† LLMs predict the next word — they don't understand meaning (Ch. 5, §5.1): The threshold concept that large language models generate text by predicting the most probable next token based on statistical patterns in training data, without any internal model of meaning, truth, or the world. See also: stochastic parrot.

M

Machine learning (Ch. 1, §1.2; Ch. 3): A subset of AI in which systems learn from data rather than being explicitly programmed with rules. The three main types are supervised learning, unsupervised learning, and reinforcement learning.

Mastery-based learning (Ch. 16, §16.6): An educational approach in which students must demonstrate mastery of a concept before advancing, with AI systems potentially enabling personalized pacing and assessment.

Measurement bias (Ch. 9, §9.2): Bias introduced when the variables used to measure a concept do not accurately capture what they purport to measure, or measure it differently for different groups. For example, using healthcare costs as a proxy for health needs introduces measurement bias because access to healthcare varies by race.

Metadata (Ch. 12, §12.2): Data about data — information such as the time a message was sent, the location from which it was sent, the duration of a call, or the size of a file. Though metadata does not contain the content of communications, it can reveal detailed patterns about behavior and relationships.

Model (Ch. 1, §1.2; Ch. 3, §3.4): In machine learning, the mathematical structure that has been trained on data and can make predictions or classifications on new inputs. A model is the product of the training process.

Model efficiency (Ch. 18, §18.5): Techniques for achieving the same or similar AI performance with less computation, including model distillation, pruning, quantization, and more efficient architectures. See also: green AI.

Multimodal AI (Ch. 21, §21.1): AI systems that can process and generate content across multiple types of input and output — text, images, audio, video — within a single model.

N

Narrow AI (weak AI) (Ch. 1, §1.3): An AI system designed to perform a specific task or small set of related tasks. All existing AI systems are narrow AI. A chess engine, a spam filter, and a large language model are all examples of narrow AI — each excels within its domain but cannot generalize beyond it. See also: artificial general intelligence.

Neural network (Ch. 1, §1.2; Ch. 3, §3.6): A computing architecture inspired by biological neurons, consisting of interconnected layers of nodes (neurons) that process information. Deep neural networks have many layers and are the foundation of deep learning.

Next-token prediction (Ch. 5, §5.1): The core mechanism of large language models: given a sequence of tokens, the model predicts the probability distribution over possible next tokens and selects one. This is how LLMs generate text, one token at a time.

O

Overfitting (Ch. 3, §3.5): A failure mode in which a model learns the noise and idiosyncrasies of its training data rather than the underlying patterns, resulting in excellent performance on training data but poor performance on new data. Like a student who memorizes answers to practice exams but cannot solve new problems. See also: underfitting, generalization.

P

Pacing problem (Ch. 13, §13.1): The fundamental challenge of regulating technology: innovation moves faster than regulation, and by the time a law is passed, the technology it addresses may have already changed.

Panopticon effect (Ch. 12, §12.4): Named after Jeremy Bentham's panopticon prison design, the psychological phenomenon in which people modify their behavior when they believe they might be observed, even when they cannot confirm whether observation is actually occurring. See also: chilling effect.

Parameter (Ch. 3, §3.6): A numerical value within a machine learning model that is adjusted during training. Large language models have billions of parameters. The parameters collectively encode the patterns the model has learned from its training data.

Personalized learning (Ch. 16, §16.3): An educational approach that tailors instruction to individual student needs, pace, and preferences. AI-powered personalized learning uses data about student performance to adapt content dynamically. See also: adaptive learning.

Pixel (Ch. 6, §6.1): The smallest addressable element of a digital image. Each pixel stores color information (typically as red, green, and blue values between 0 and 255).

Power usage effectiveness (PUE) (Ch. 18, §18.1): A metric for data center energy efficiency: the ratio of total facility energy to IT equipment energy. A PUE of 1.0 would mean all energy goes to computing; typical data centers have PUE values of 1.2–1.6.

Precautionary principle (Ch. 13, §13.5): The regulatory philosophy that when an action risks causing harm to the public or environment, precautionary measures should be taken even if the causal relationship is not fully established. Contrasted with the innovation principle.

Predictive policing (Ch. 17, §17.1): AI systems that use historical crime data to predict where crimes are likely to occur, used to guide police patrol deployment. Criticized for reinforcing feedback loops and amplifying historical policing biases.

Pre-training (Ch. 5, §5.3): The initial phase of training a large language model, in which the model learns language patterns by predicting the next token on massive amounts of text data. See also: fine-tuning, RLHF.

Prompt (Ch. 14, §14.1): The text input a user provides to a language model to elicit a response. The quality and structure of the prompt significantly affect the quality of the output.

Prompt engineering (Ch. 14, §14.2): The practice of crafting prompts to elicit better or more specific responses from language models. Includes techniques like zero-shot, few-shot, and chain-of-thought prompting.

Protected class (Ch. 9, §9.4): A group of people sharing a characteristic (such as race, gender, age, disability, or religion) that is legally protected from discrimination.

Proxy variable (Ch. 9, §9.2): A variable that is not itself a protected characteristic but is correlated with one and can therefore serve as a stand-in for discriminatory decision-making. Zip code, for example, can be a proxy for race in many U.S. cities.

R

Rebound effect (Jevons paradox) (Ch. 18, §18.6): The phenomenon in which efficiency improvements lead to increased total consumption rather than reduced consumption. If AI makes something cheaper or easier, people may use more of it, potentially offsetting environmental gains.

Recidivism prediction (Ch. 17, §17.2): AI systems that estimate the likelihood that a criminal defendant or prisoner will reoffend. Used to inform bail, sentencing, and parole decisions. See also: risk assessment instrument.

Red-teaming (Ch. 20, §20.4): The practice of having a team attempt to find vulnerabilities, failures, and harmful behaviors in an AI system before deployment, simulating adversarial or misuse scenarios.

Regulatory capture (Ch. 13, §13.5): A phenomenon in which a regulatory agency becomes dominated by the industry it is supposed to regulate, leading to regulations that serve industry interests rather than public interests.

Regulatory sandbox (Ch. 13, §13.2): A controlled environment in which companies can test new AI technologies under relaxed regulatory requirements, allowing regulators to learn about the technology while providing limited, supervised deployment.

Reinforcement learning (Ch. 3, §3.3): A machine learning paradigm in which an agent learns to make decisions by taking actions in an environment and receiving rewards or penalties based on outcomes. The approach behind game-playing AI (AlphaGo) and RLHF. See also: reward hacking.

Reinforcement learning from human feedback (RLHF) (Ch. 5, §5.4; Ch. 20, §20.4): A training technique in which a language model is fine-tuned using feedback from human evaluators who rank the quality of model outputs. The primary method for making LLMs "helpful, harmless, and honest."

Representation bias (Ch. 9, §9.2): Bias that occurs when the training data does not adequately represent the population the AI system will serve — for example, a facial recognition system trained predominantly on lighter-skinned faces.

Reward hacking (Ch. 20, §20.2): A failure mode in which an AI agent finds unintended ways to maximize its reward signal that do not align with the designer's actual goals. See also: specification gaming, alignment problem.

Right to be forgotten (Ch. 12, §12.5): Under GDPR, an individual's right to request that organizations delete their personal data under certain circumstances.

Right to explanation (Ch. 17, §17.3): The principle (codified in some jurisdictions, notably the EU) that individuals have a right to a meaningful explanation of automated decisions that significantly affect them.

Risk assessment instrument (RAI) (Ch. 17, §17.2): A tool — increasingly AI-powered — used in the criminal justice system to estimate a defendant's risk of reoffending or failing to appear in court. See also: COMPAS, recidivism prediction.

Risk-based regulation (Ch. 13, §13.2): A regulatory approach that classifies AI systems by the level of risk they pose and imposes proportional requirements. The EU AI Act uses a four-tier risk classification: unacceptable, high, limited, and minimal risk.

Robustness (Ch. 20, §20.2): An AI system's ability to maintain reliable performance under varied, unexpected, or adversarial conditions. See also: adversarial example, distributional shift.

S

Scenario planning (Ch. 21, §21.3): A strategic thinking technique that develops multiple plausible future scenarios (optimistic, realistic, cautionary) to prepare for uncertainty rather than relying on a single prediction.

Scope 1/2/3 emissions (Ch. 18, §18.1): A framework for categorizing greenhouse gas emissions. Scope 1: direct emissions from owned sources. Scope 2: indirect emissions from purchased electricity. Scope 3: all other indirect emissions in the value chain. For AI, most emissions are Scope 2 (data center electricity) and Scope 3 (hardware manufacturing, cooling water).

Selection bias (Ch. 4, §4.4): Bias introduced when the data collected is not representative of the population the AI system will serve, typically because the data collection process systematically excludes certain groups.

Self-regulation (Ch. 13, §13.5): The practice of an industry establishing and enforcing its own standards and guidelines, without legally binding government regulation. See also: soft law.

Sociotechnical system (Ch. 21, §21.5): A framework for understanding AI that recognizes it as embedded in social, institutional, and political contexts — not as a purely technical artifact. AI systems are shaped by and shape the societies that create and use them.

Soft law (Ch. 13, §13.5): Non-binding guidelines, principles, and standards that influence behavior through reputation, industry norms, and voluntary compliance rather than legal enforcement. Many AI governance frameworks are currently soft law. See also: hard law.

Specification gaming (Ch. 20, §20.2): When an AI system satisfies the literal specification of its objective while violating the spirit of what the designer intended. A form of reward hacking.

Stochastic parrot (Ch. 5, §5.6): A term coined in the 2021 Bender et al. paper (see Appendix B) to describe large language models as systems that generate plausible-sounding text by recombining patterns from training data without understanding meaning. The term remains contested.

Style transfer (Ch. 11, §11.1): An AI technique that applies the visual style of one image (such as a Van Gogh painting) to the content of another image (such as a photograph).

Supervised learning (Ch. 3, §3.1): A machine learning paradigm in which the model learns from labeled examples — input-output pairs where the correct answer is provided. The most common type of machine learning.

Superintelligence (Ch. 20, §20.3): A hypothetical AI system that vastly exceeds human cognitive abilities across all domains. A central concept in long-term AI safety discussions.

Surveillance capitalism (Ch. 12, §12.1): A term coined by Shoshana Zuboff describing an economic system in which the collection, analysis, and sale of personal data for behavioral prediction and modification is the primary profit driver. See also: data broker.

T

Task-based framework (Ch. 10, §10.2): An approach to analyzing AI's impact on work that focuses on individual tasks within jobs rather than whole jobs, recognizing that AI typically automates specific tasks rather than entire occupations.

Techno-nationalism (Ch. 19, §19.2): The strategic pursuit of AI dominance as a component of national power and security, motivating government investment, export controls, and competition between nations.

Technological determinism (Ch. 21, §21.5): The belief that technology develops autonomously and determines social outcomes, leaving humans as passive recipients. Contrasted with the view that humans shape technology through choices, policies, and values.

Test set (Ch. 3, §3.4): A portion of data held back from training and validation, used to evaluate the model's final performance on data it has never seen. See also: training set, validation set.

Training data (Ch. 1, §1.2; Ch. 3, §3.4): The data used to train a machine learning model — the examples from which the model learns patterns. The quality, representativeness, and biases of training data directly determine the model's behavior.

Training set (Ch. 3, §3.4): The portion of data used to train the model — the examples the model learns from. See also: validation set, test set.

Training vs. inference energy (Ch. 18, §18.1): The distinction between the energy consumed during the one-time process of training a model (which can be enormous for large models) and the ongoing energy consumed each time the model processes a query (inference), which is smaller per query but adds up at scale.

Transfer learning (Ch. 5, §5.4): Training a model on one task or domain and then adapting it to a different but related task. Pre-training a language model on general text and then fine-tuning it for medical questions is an example.

Transformer (Ch. 2, §2.5; Ch. 5, §5.2): A neural network architecture introduced in the 2017 paper "Attention Is All You Need" (Vaswani et al.) that processes sequences using self-attention mechanisms. The foundation of all large language models. See also: self-attention.

Turing Test (Ch. 2, §2.1): A test proposed by Alan Turing in 1950: if a human evaluator, communicating via text, cannot reliably distinguish between a machine and a human, the machine is said to have passed the test. Influential historically but limited as a definition of intelligence.

U

Underfitting (Ch. 3, §3.5): A failure mode in which a model is too simple to capture the underlying patterns in the data, resulting in poor performance on both training data and new data. Like a student who only learned the chapter titles and cannot answer any detailed questions. See also: overfitting.

Unsupervised learning (Ch. 3, §3.2): A machine learning paradigm in which the model identifies patterns and structures in data without labeled examples. Used for clustering, dimensionality reduction, and anomaly detection.

V

Validation set (Ch. 3, §3.4): A portion of data used during training to tune the model's configuration and check for overfitting, separate from both the training set and the test set.

Value alignment (Ch. 20, §20.1): The challenge of ensuring that an AI system's goals and behaviors are aligned with human values. See also: alignment problem, constitutional AI.

Verification (Ch. 8, §8.6): The process of independently checking an AI system's outputs for accuracy, typically using external sources, domain expertise, or cross-referencing. See also: hallucination verification.

Z

Zero-shot prompting (Ch. 14, §14.2): A prompting technique in which the user gives a language model a task without providing any examples, relying on the model's pre-training to understand and perform the task. See also: few-shot prompting, chain-of-thought prompting.