Glossary

AI governance glossary

399 terms from AI law, risk management and the IAPP AIGP Body of Knowledge, each defined in a sentence or two. Terms marked Full explainer have their own page with an example and the exam angle.

Written with AI from the sources below and checked against our facts sheet. Last checked 7 Nov 2026. How this site is made · Spot an error?

A

Acceptable use policy (AUP): Rules binding employees and others on how they may use AI tools: approved tools, prohibited inputs and uses, review and reporting. Lesson I.C.5
Accountability: Identifiable people and organizations answer for an AI system's outcomes and have the authority to act. Lesson I.A.4
Adverse action notice: A required notice telling a consumer about a negative decision and, depending on the law, the reasons or the consumer report behind it. Lesson II.B.3
Agentic AI: AI that plans and takes multi-step actions, often using tools, to pursue a goal with limited human intervention. Lesson I.A.1
Agentic architecture: A design in which a model plans steps and calls tools to take actions toward a goal with limited human input. Lesson IV.A.4
AGPL network clause: The GNU Affero GPL term that treats offering modified code as a network service like distributing it, triggering source-sharing duties. Lesson IV.B.4
AI actor: Anyone who plays an active role in the AI system life cycle, including organizations and individuals that deploy or operate AI. Lesson II.D.1
AI bill of materials: An inventory of the models, datasets and components that make up an AI product. Lesson I.C.4
AI business operator: Under Korea's Act, an entity that develops AI or uses AI to provide products or services. Lesson II.C.12
AI disclosure: Telling people, at or before first interaction, that they are dealing with an AI system or AI-generated content. Lesson IV.C.6
AI governance: AI governance is the set of policies, roles, processes and controls an organization uses to direct how AI is built, bought and used. It assigns accountability, sets risk appetite and keeps each AI system lawful, safe, fair and aligned with the organization's values throughout its life. Full explainer · Lesson I.B.1
AI governance committee: A cross-functional body that approves AI policy, reviews higher-risk use cases and resolves escalations. Lesson I.B.1
AI impact assessment: A structured review, done before build and revisited over time, of how an AI system could affect people, groups and society. Lesson III.A.2
AI incident: An event where an AI system's development, use or malfunction causes or nearly causes harm. Lesson I.C.2
AI incident database: A public collection of real-world AI harms and near misses (e.g., the AI Incident Database, OECD AIM) used for learning, not regulatory reporting. Lesson III.C.7
AI Incidents Monitor (AIM): An OECD database tracking AI incidents and hazards reported in the news, to build evidence for policy. Lesson II.D.1
AI inventory: An organization-wide register of AI systems with their owners, purposes, risk tiers and status. Lesson I.C.2
AI literacy: AI literacy is the skills, knowledge and understanding needed to use AI in an informed way and to recognize its opportunities, risks and harms. Under Art. 4 of the EU AI Act, providers and deployers must take measures to support the development of AI literacy of their staff. This covers every AI system, not only high-risk ones. Full explainer · Lesson I.B.3
AI Office: The European Commission's AI body that supervises GPAI models and, after the Omnibus, certain AI systems built on them; it can investigate and fine GPAI providers. Lesson II.C.9
AI regulatory sandbox: A controlled framework set up by a competent authority where providers can develop and test AI under regulatory supervision. Lesson II.C.9
AI risk management: AI risk management is the ongoing process of identifying, assessing, ranking and treating the risks an AI system poses to people, organizations and society. It weighs each harm's probability and severity against the organization's risk tolerance. Monitoring continues across the whole life cycle, not just before launch. Full explainer · Lesson III.A.5
AI RMF Playbook: NIST's voluntary online companion offering suggested actions, documentation prompts and references for every subcategory. Lesson II.D.3
AI system (OECD / EU AI Act): A machine-based system that infers from inputs how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments. Lesson I.A.1
AI system inventory: An organization-wide register of AI systems with their owners, documentation, incident plans, data dictionaries and links to code. Lesson III.C.2
AI system owner: The business leader accountable for a specific AI system's purpose, performance and compliance. Lesson I.B.1
AI washing: Overstating or misrepresenting the use or capabilities of AI in a product. Lesson II.B.4
Alert threshold: A pre-set value that, when crossed, signals a problem and triggers action by a named owner. Lesson III.C.3
Algorithmic disgorgement: An FTC remedy requiring deletion of models and algorithms built from improperly obtained data. Lesson II.B.4
Annex I: The list of EU product-safety laws (medical devices, toys, machinery, vehicles and more) whose regulated products can make embedded AI high-risk. Lesson II.C.11
Annex I route: AI that is, or is a safety component of, a product covered by listed EU product-safety laws and requiring third-party conformity assessment (Art. 6(1)). Lesson II.C.3
Annex III: The list of eight sensitive areas of use that make an AI system high-risk, such as employment, credit and education. Lesson II.C.11
Annex III route: AI used in one of eight listed sensitive areas, such as employment, credit, education or law enforcement (Art. 6(2)). Lesson II.C.3
Annex IV: The required contents of a high-risk system's technical documentation. Lesson II.C.11
Art. 6(3) derogation: The exit from Annex III high-risk status for systems posing no significant risk, such as narrow procedural or preparatory tasks; never available if the system profiles people. Lesson II.C.3
Article 22: The GDPR right not to be subject to solely automated decisions with legal or similarly significant effects, subject to three exceptions and safeguards. Lesson II.A.4
Artifact signing: Cryptographically signing model files so the pipeline can verify they haven't been altered before deployment. Lesson III.C.10
Artificial Intelligence and Data Act (AIDA): Canada's proposed federal AI law, part of Bill C-27, which died when Parliament was prorogued in January 2025. Lesson II.C.15
Audit: An independent or internal check of a system or program against defined criteria such as laws, standards, policies or the provider's own claims. Lesson III.C.5
Audit rights: The customer's contractual right to obtain evidence of, or directly check, the vendor's controls and compliance. Lesson IV.B.2
Authorized representative: An EU-established entity with a written mandate from a non-EU provider to carry out specified obligations on its behalf. Lesson II.C.7
Automated decision-making technology (ADMT): Technology that processes personal data to produce outputs, such as scores or recommendations, used to make or assist decisions about people. Lesson II.C.14
Automated decisionmaking technology (ADMT): California's term for technology that replaces or substantially replaces human decision-making, regulated under its 2026 privacy rules. Lesson II.A.4
Automation bias: The tendency of people to over-rely on automated outputs and stop exercising their own judgment. Lesson I.A.5
Awareness program: Ongoing communication that keeps AI risks, rules and reporting channels front of mind for everyone. Lesson I.B.3

B

Backdoor: A hidden behavior planted through poisoning that activates only when a specific trigger appears. Lesson III.B.5
Baseline: How the task performs today without the AI, used to judge whether the system is actually an improvement. Lesson III.A.1
Benchmarking: Comparing a system's performance offline against a reference such as the current process, another model or a standard test set. Lesson III.A.6
Bias audit (NYC LL 144): An annual, independent evaluation of an automated employment decision tool's impact by sex and race/ethnicity, with a public summary. Lesson II.C.14
Biometric data: Personal data from technical processing of physical, physiological or behavioral traits that allows or confirms a person's unique identification. Lesson II.A.5
BIPA: Illinois' Biometric Information Privacy Act, requiring written notice and release before collecting biometrics and allowing private lawsuits. Lesson II.A.5
Brittleness: A system's tendency to fail, often abruptly, when conditions differ from those it was trained on. Lesson III.C.8
Bug bash vs. bug bounty: A bug bash is a focused internal session where builders, testers and sometimes users hunt functional bugs together; a bug bounty pays outside researchers to report flaws responsibly. Lesson III.C.5
Build vs. buy: The choice between developing your own AI model and licensing one, which shifts control, obligations and liability. Lesson IV.B.3
Build, buy or adapt: The choice between training your own model, buying a vendor system and fine-tuning an existing model. Lesson III.A.3
Business case: The argument for an AI project: the problem, why AI beats simpler options, measurable objectives, and direct and indirect costs weighed against benefits. Lesson III.A.1
Business necessity: The US employer defense that a practice causing disparate impact is job-related and necessary, defeated if a less discriminatory alternative was refused. Lesson II.B.2

C

Canary (phased) release: Releasing to a small share of users first and expanding only if monitoring stays healthy. Lesson III.C.1
Catalogue of Tools & Metrics for Trustworthy AI: An OECD.AI library of tools and metrics that help put the principles into practice. Lesson II.D.1
Catastrophic forgetting: When retraining or fine-tuning on new data makes a model abruptly lose capabilities it had before. Lesson III.C.8
Category and subcategory: The outcome statements each AI RMF function breaks into: 19 categories and 72 subcategories in total. Lesson II.D.3
CE marking: The mark a provider affixes to show a high-risk AI system conforms to the AI Act's requirements (Art. 48). Lesson II.C.5
Centralized governance: One central function sets AI policy and makes AI decisions for the whole organization. Lesson I.B.4
Certificate of destruction: Written confirmation from a provider or contractor that specified data or model copies have been permanently destroyed. Lesson IV.C.8
Chain of custody: An unbroken record of who handled something and what they did, borrowed from evidence handling. Lesson III.B.2
Champion–challenger: Testing a candidate model against the production model on live cases and promoting it only if it clearly performs better. Lesson III.C.4
Change management: The controlled process for updating a system: test, version, document, approve and release with rollback ready. Lesson III.C.4
Change-notification clause: A contract term requiring the vendor to tell you before material model updates or new AI features go live. Lesson I.C.4
Chapter V: The part of the AI Act (Arts. 51–56) setting obligations for general-purpose AI models. Lesson II.C.11
Classic (predictive) AI: Models that classify, score or forecast from inputs rather than create new content. Lesson IV.A.2
Cloud deployment: Running an AI system on a provider's remote infrastructure, typically as a managed service. Lesson IV.A.3
Compatibility assessment: The Article 6(4) analysis of link, context, nature of data, consequences and safeguards used to decide whether a new purpose fits the original one. Lesson II.A.1
Concept drift: A change in the relationship between inputs and the correct output, so old patterns stop holding. Lesson I.A.3
Confabulation: Generative AI confidently producing false or fabricated content; one of the 12 GenAI Profile risks. Lesson II.D.3
Confirmation bias: The human tendency to favor information that supports existing beliefs. Lesson III.B.4
Conformity assessment: The process of showing a high-risk system meets the requirements, either by internal control or with a notified body (Art. 43). Lesson II.C.5
Consequential decision: A decision with significant effects on access to things like employment, housing, credit, insurance, health care or education. Lesson II.C.14
Containment: The response step that stops harm spreading, for example by disabling a feature or rolling back, while preserving evidence. Lesson III.C.7
Continuous monitoring: Ongoing tracking of a deployed AI system's performance, inputs, outputs, fairness, security and human feedback. Lesson III.C.3
Contributing factor: A condition, such as a rushed timeline or skills gap, that makes harm more likely or more severe. Lesson III.A.5
Controller: The party that decides the purposes and essential means of processing personal data, and carries primary accountability. Lesson II.A.3
Copyleft license: An open license, such as the GPL, that requires derivative works to be released under the same terms when they are distributed. Lesson IV.B.4
Counterfactual explanation: An explanation that states the smallest change to the inputs that would have changed the outcome. Lesson III.B.7
Covered platform: Under TAKE IT DOWN, a public website, online service or app that primarily hosts user-generated content. Lesson II.C.13
Cross-functional collaboration: Bringing legal, privacy, security, technical, business, HR and other functions together to govern AI jointly. Lesson I.B.2

D

Data availability: Whether suitable, good-quality data exists and the organization has the right to use it for the intended purpose. Lesson IV.A.1
Data drift: A change in the input data a deployed model receives compared with its training data. Lesson I.A.3
Data integrity: Assurance that data has stayed complete and unaltered, protected from corruption and tampering. Lesson III.B.1
Data leakage: When a training feature carries information about the outcome that won't be available at prediction time, inflating scores. Lesson III.B.3
Data lineage: The traceable path of data from its source through every transformation into a model. Lesson I.C.3
Data minimization: Processing only personal data that is adequate, relevant and limited to what is necessary for the stated purpose. Lesson II.A.2
Data owner vs. data steward: The owner is the accountable leader who decides how a dataset may be used; the steward manages its quality, documentation and day-to-day handling. Lesson III.B.1
Data poisoning: An attack that corrupts training data so the model learns harmful or wrong behavior. Lesson I.C.3
Data protection by design and by default: GDPR Article 25's duty to build in measures that make the principles effective, and to process only what each purpose needs by default. Lesson II.A.2
Data protection impact assessment (DPIA): A GDPR-mandated assessment, done before processing, of high-risk processing's necessity, risks and mitigations. Lesson II.A.3
Data provenance: A documented record of where data came from and how it has been transformed. Lesson I.C.1
Data residency: A requirement that data be stored or processed within a particular country or region. Lesson IV.A.3
Dataset versioning: Keeping identifiable versions of datasets so each model can be linked to the exact data it used. Lesson III.B.2
Datasheet for a dataset: A standard document describing a dataset's motivation, composition, collection, processing, uses, distribution and maintenance. Lesson III.B.6
Datasheet for datasets: A document recording a dataset's motivation, composition, collection process and recommended uses. Lesson III.C.2
Deactivation policy: A pre-agreed policy setting the triggers, decision rights, controls, fallbacks, notifications and restoration criteria for switching an AI system off or restricting it. Lesson IV.C.7
Decentralized governance: Each business unit governs its own AI with little central coordination. Lesson I.B.4
Deceptive practice: A material representation or omission likely to mislead a consumer acting reasonably, judged by its net impression. Lesson II.B.4
Decommissioning: The controlled process of retiring a system: mapping dependencies, notifying users, transitioning, archiving or destroying data and models, and closing the records. Lesson IV.C.8
Deepfake: AI-generated or manipulated image, audio or video that resembles real people, places or events and would falsely appear authentic. Lesson II.C.3
Defense in depth: Layering safeguards at different points in a system so a single failure doesn't cause harm. Lesson III.B.8
Deployer: An organization that uses an AI system under its own authority, outside purely personal use. Lesson I.B.5
Deployer impact assessment: An assessment of how a selected AI system will affect people and the organization in the deployer's own context, done before use and kept up to date. Lesson IV.B.1
Design defect: A flaw in the design itself that makes every unit unreasonably dangerous, typically where a reasonable safer alternative existed. Lesson II.B.5
Deterministic output: A result that is always the same for the same input, as with conventional rule-based software. Lesson I.A.3
Developer: The party that designs, builds, trains and tests an AI model or system. Lesson I.B.5
Differential privacy: Adding calibrated statistical noise so individuals can't be singled out while overall patterns survive. Lesson III.B.5
Digital Omnibus on AI: Regulation (EU) 2026/1744, in force 27 July 2026, which deferred high-risk dates, added two prohibitions and simplified parts of the AI Act. Lesson II.C.10
Disaggregated evaluation: Breaking performance and error rates down by relevant groups to reveal problems that overall averages hide. Lesson IV.B.1
Disparate impact: A facially neutral practice that disproportionately disadvantages a protected group, regardless of intent. Lesson II.B.2
Disparate treatment: Intentionally treating people differently because of a protected characteristic, including deliberately using a proxy for it. Lesson II.B.2
Distributor: Any supply-chain actor, other than the provider or importer, that makes an AI system available on the EU market. Lesson II.C.7
Diversity of expertise and perspective: Including different disciplines and backgrounds so the group spots risks a single viewpoint would miss. Lesson I.B.2
Domestic representative: A local agent that foreign AI operators above set thresholds must appoint to deal with Korean authorities. Lesson II.C.12
Downstream harm: Harm that ripples beyond the original decision or user, including to people who never used the system. Lesson IV.C.5
DPIA: GDPR's data protection impact assessment, required of controllers before personal-data processing likely to result in high risk. Lesson III.A.2
Draft vs. final text: A legislative proposal or negotiating draft can change before adoption; only the final published text binds, so figures quoted from drafts go stale. Lesson II.C.15

E

ECOA and Regulation B: The US credit anti-discrimination law and its implementing rule, which also require specific reasons when credit is denied. Lesson II.B.3
Edge deployment: Running an AI model directly on an end device such as a phone, camera, vehicle or sensor. Lesson IV.A.3
Emotion recognition system: AI that identifies or infers people's emotions or intentions from biometric data such as facial expressions or voice. Lesson II.C.2
Entry into force vs application: A law can be in force (legally existing) before its obligations apply; the AI Act entered into force on 1 Aug 2024 but applies in stages. Lesson II.C.10
Environment separation: Keeping development, test or staging, and production apart, with models promoted only through an approved pipeline. Lesson III.C.10
EO 14365: The December 2025 executive order seeking a national AI policy framework and creating a DOJ task force to challenge state AI laws. Lesson II.C.13
Ethics by design: Identifying and resolving ethical issues from planning and design onward, so values shape the system rather than being checked at the end. Lesson I.C.1
EU AI Act Art. 9(5) order: For high-risk providers: eliminate or reduce risks by design, then add mitigation and control measures, then give deployers information and, where appropriate, training. Lesson III.A.7
EU database for high-risk AI: The public EU register where providers record high-risk systems, including electronic instructions for use for most of them (Arts. 49 and 71). Lesson III.C.9
European AI Board: A body of Member State representatives that advises the Commission and promotes consistent application of the AI Act; it has no fining power. Lesson II.C.9
Excessive agency: Giving an AI agent more permissions, tools or autonomy than its task needs, so a manipulation can do more damage. Lesson III.C.6
Executive champion: A senior leader who sponsors AI governance and gives it visible backing, authority and budget. Lesson I.B.1
Exit assistance: Contract terms covering data return, certified deletion and transition support when the relationship ends. Lesson IV.B.2
Expert system: A knowledge-based system that reasons over facts and if-then rules supplied by human experts to reach new conclusions; it can count as AI because it infers from encoded knowledge. Lesson I.A.1
Explainable AI: Explainable AI (XAI) is AI that lets people understand how it reached a particular output, such as the factors behind a loan denial. It covers models that are interpretable by design and after-the-fact methods such as SHAP, LIME and counterfactual explanations. Each explanation should suit the audience that needs it. Full explainer · Lesson I.A.4
Explicit consent: A clear, express statement of consent (for example, a signed or separately confirmed statement), required for some Article 9 and Article 22 exceptions. Lesson II.A.5
External communication plan: A pre-agreed plan for what an organization tells users, affected people, regulators, partners and the public about its AI, when and through which channels. Lesson IV.C.6
Extraterritorial scope: The AI Act's reach to non-EU providers and deployers when their AI is placed on the EU market or its output is used in the EU. Lesson II.C.1

F

F1 score: A single score that balances precision and recall; it falls sharply if either one is poor. Lesson III.B.7
Failure to warn: A defect arising from inadequate instructions or warnings about foreseeable risks. Lesson II.B.5
Fair Credit Reporting Act (FCRA): The US law governing consumer reports, with notice and dispute rights when a report drives an adverse decision. Lesson II.B.3
Fair Housing Act: The 1968 US law banning discrimination in housing sales, rentals, lending and advertising, including through disparate impact. Lesson II.B.3
Fair use: A US defense that permits some unlicensed use of copyrighted works, decided by weighing purpose, nature, amount and market effect. Lesson II.B.1
Fairness: Ensuring AI outcomes don't unjustly disadvantage individuals or groups, with harmful bias identified and managed. Lesson I.A.4
Fairness metrics: Yardsticks such as demographic parity, equal opportunity and calibration that compare outcomes across groups; when base rates differ, they can't all be met at once. Lesson III.B.7
False positive / false negative: A false positive flags something that isn't there (a false alarm); a false negative misses something that is. Lesson III.B.7
Feature flag: A switch that turns a feature on or off for a jurisdiction, customer group or cohort without redeploying software. Lesson IV.C.7
Federated learning: Training a model across separate data holders so raw data stays put and only model updates are shared. Lesson III.B.5
Feedback loop: When a model's outputs influence the data it is later trained or tested on, reinforcing its own patterns. Lesson III.B.8
Fine-tuning: Further training a pre-trained model on your own examples, which permanently changes its weights. Lesson IV.A.4
Fishbone (Ishikawa) diagram: A root-cause tool that groups candidate causes into categories such as data, model, process, people and environment. Lesson III.C.8
Fitness for purpose: Whether data suits the specific task and population a model will serve, beyond being large and accurate. Lesson III.B.1
Foundation model: A large model trained on broad data that can be adapted to many downstream tasks. Lesson I.A.1
Four Ts: A business risk-response menu: treat, transfer, tolerate or terminate. Lesson III.A.7
Four-fifths rule: A rule of thumb from the 1978 Uniform Guidelines: a selection rate below 80% of the highest group's rate signals possible adverse impact. Lesson II.B.2
Fourth party: Your vendor's own supplier, such as the company whose foundation model the vendor builds on. Lesson I.C.4
Framework for the Classification of AI Systems: An OECD tool that profiles AI systems across five dimensions to connect their characteristics with policy and risk considerations. Lesson II.D.1
FRIA: The EU AI Act's fundamental rights impact assessment, a duty on certain deployers of high-risk systems before first use. Lesson III.A.2
Frontier developer (SB 53): A developer that trained a foundation model using more than 10²⁶ operations; 'large' frontier developers also exceed $500M in annual revenue. Lesson II.C.14
FTC Act Section 5: The US federal prohibition on unfair or deceptive acts or practices in commerce, applied to AI without any AI-specific statute. Lesson II.B.4
Function creep: The gradual expansion of a system into tasks it was never designed or assessed for. Lesson IV.C.5
Fundamental rights: Rights protected by the EU Charter, such as dignity, privacy, nondiscrimination and a fair trial, which the AI Act aims to safeguard alongside health and safety. Lesson II.C.1
Fundamental rights impact assessment: A fundamental rights impact assessment (FRIA) is an EU AI Act duty (Art. 27) for certain deployers of high-risk AI, completed before first use. It records how the system will be used, who it affects and the specific risks of harm to them. It also covers human oversight and the response if risks materialize. Full explainer · Lesson II.C.6

G

Gap analysis: Reviewing existing policies against AI's characteristics to find what they fail to address. Lesson I.C.3
General-purpose AI model: A general-purpose AI model is an AI model, such as a large language model, that can competently perform a wide range of distinct tasks. It shows significant generality and can be built into many downstream systems. Under the EU AI Act (Art. 3(63)), its provider has its own obligations, which have applied since 2 August 2025. Full explainer · Lesson II.C.8
Generative AI: Models that create new content such as text, images, code or audio. Lesson IV.A.2
Go/no-go criteria: Acceptance thresholds agreed in advance (performance, fairness, robustness, security, documentation) that a system must meet before release. Lesson III.C.1
GOVERN: The cross-cutting function that builds the policies, accountability, culture and processes the other functions rely on. Lesson II.D.2
GPAI Code of Practice: A voluntary code (July 2025) with transparency, copyright, and safety and security chapters that providers can rely on to demonstrate compliance. Lesson II.C.8
GPAI system: An AI system based on a general-purpose AI model, able to serve a variety of purposes directly or when integrated into other systems. Lesson II.C.8
Graceful degradation: Falling back to human handling or a simpler process so service continues when the AI is switched off. Lesson IV.C.7
Ground truth: The verified correct outcome that predictions are compared against to measure real accuracy. Lesson III.C.3
Group harm: Damage to a community or population sub-group, such as systematic discrimination or targeted surveillance. Lesson I.A.2

H

Hallucination: Fluent, confident output from a generative model that is false or fabricated. Lesson I.A.3
Harms matrix: A probability × severity grid used to rank potential harms so the worst get attention first. Lesson III.A.5
High-impact AI (Korea): AI that may significantly affect human life, physical safety or fundamental rights in listed areas such as health care, hiring, lending and public services. Lesson II.C.12
High-performance AI (Korea): AI trained with cumulative compute of 10²⁶ FLOPs or more, which carries life-cycle safety duties. Lesson II.C.12
High-risk AI system: A high-risk AI system is an AI system the EU AI Act allows but subjects to its strictest rules. Either it is a safety component of a regulated product needing third-party assessment (Art. 6(1), Annex I). Or it is used in one of eight sensitive Annex III areas, such as hiring or credit scoring. Full explainer · Lesson II.C.3
Historical bias: Bias from data that accurately records a past shaped by unfair treatment. Lesson III.B.4
Holding statement: A pre-drafted initial public statement used in the first hours of an incident while facts are confirmed. Lesson IV.C.6
Human authorship: The US requirement that a work be created by a human to qualify for copyright protection. Lesson II.B.1
Human oversight: Human oversight is people's ability to monitor an AI system, understand its limits, interpret its outputs and override or stop it. Under Art. 14 of the EU AI Act, providers must design high-risk systems so trained people can oversee them effectively. Deployers must assign those people (Art. 26(2)). Full explainer · Lesson III.A.4
Human-centric AI: AI designed to augment people, respect their rights and keep meaningful human control over consequential decisions. Lesson I.A.4
Human-in-command: Oversight in which people decide whether, when and how an AI system is used at all, and can overrule or switch it off. Lesson III.A.4
Human-in-the-loop: A person must review and approve each decision before it takes effect. Lesson I.A.5
Human-on-the-loop: The system acts autonomously while a person monitors it and can intervene or halt it. Lesson I.A.5
Human-out-of-the-loop: The AI operates fully automatically with no real-time human involvement. Lesson III.A.4
Hybrid (federated) governance: Central policy and oversight combined with local assessment of use cases; the usual fit for large multinationals. Lesson I.B.4
Hybrid deployment: Combining environments, such as edge or on-premise inference with cloud training and updates, so each workload runs where it fits. Lesson IV.A.3

I

IAPP planning order: Business problem → use case → applicable laws → gaps and risks → data needed. Lesson III.A.1
Implementation tiers: A NIST Cybersecurity Framework feature describing how rigorous an organization's practices are; the AI RMF 1.0 has no tiers. Lesson II.C.15
Importer: An EU-established entity that places on the EU market an AI system bearing the name or trademark of a provider established outside the EU. Lesson II.C.7
Incident record: A documented account of what happened, its impact, containment, root cause, notifications, corrective actions and lessons. Lesson IV.C.4
Indirect prompt injection: Malicious instructions hidden in content the model reads, such as a retrieved document or web page, that hijack its behavior. Lesson IV.A.4
Individual harm: Damage to a specific person's rights, safety, civil liberties or economic opportunity. Lesson I.A.2
Inference harm: A privacy harm in which an AI derives sensitive traits a person never disclosed. Lesson I.A.2
Instructions for use: Information a provider must give deployers about a high-risk system's purpose, performance, limitations, oversight and maintenance (Art. 13). Lesson II.C.4
Intended purpose: The use a provider intends for an AI system, including its context and conditions of use; it drives risk classification. Lesson II.C.1
Intent-based liability: Liability that requires proof the AI was built or used for a prohibited purpose; disparate impact alone isn't enough (the Texas model). Lesson II.C.14
Inter-rater reliability: How consistently independent labelers give the same label to the same item; low agreement signals label noise. Lesson III.B.8
Interpretability: In NIST's terms, understanding why a system produced an output and what it means in context for the user. Lesson I.A.4
Interpretable model: A model whose decision logic people can follow directly, such as a scorecard or a small decision tree. Lesson III.A.3
IP assignment: A contract clause transferring ownership of work created by employees or contractors to the company. Lesson IV.B.5
IP indemnity: A vendor's promise to defend and compensate the customer if the AI system or its outputs infringe third-party intellectual property. Lesson IV.B.2
ISO/IEC 22989: The 2022 international standard defining AI concepts and terminology, including stakeholder roles and life-cycle stages. Lesson II.D.4
ISO/IEC 23894: Guidance on AI risk management, adapting ISO 31000 to AI. Lesson II.D.4
ISO/IEC 42001 (AIMS): The 2023 certifiable standard setting requirements for an AI management system, using Plan-Do-Check-Act. Lesson II.D.4
ISO/IEC 42005: The 2025 guidance standard on assessing an AI system's impacts on individuals, groups and society. Lesson II.D.4
ISO/IEC Guide 51 three-step method: The safety-standards order for reducing risk: inherently safe design, then protective measures, then information for use. Lesson III.A.7
Issue management: The process for capturing, triaging, escalating and resolving problems with a live system, and learning from them. Lesson IV.C.1

J

Jailbreak: A technique that coaxes a model into bypassing its safety rules or usage restrictions. Lesson III.C.6

K

Kill switch: A control that stops an AI system globally and quickly, ideally leaving it in a safe state. Lesson IV.C.7

L

Lawful basis: One of the six GDPR grounds (consent, contract, legal obligation, vital interests, public task, legitimate interests) that must apply before personal data is processed. Lesson II.A.1
Lawful rights to data: The documented legal basis, consent, license or contract that permits an organization to collect and use a dataset for a specific purpose. Lesson III.B.1
Legacy system: An AI system or model already on the market before the relevant application date, which may benefit from transitional rules. Lesson II.C.10
Legitimate interests assessment (LIA): A documented three-step test of purpose, necessity and balancing that a controller runs before relying on legitimate interests. Lesson II.A.1
Liability cap: A contractual ceiling on the damages one party can recover from the other, often with carve-outs for specific harms. Lesson IV.B.2
Liability reform: Changing the legal rules that decide who is accountable for harm, as opposed to cost-shifting tools like insurance or recalls. Lesson II.B.5
Local vs. global explanation: A local explanation shows why one output occurred; a global explanation shows what drives the model's behavior overall. Lesson IV.B.5
Localization: Restricting or adapting how an AI system operates by region, for example disabling it in one jurisdiction or running a country-specific model. Lesson IV.C.7

M

Machine learning: AI that learns patterns from data rather than being explicitly programmed with rules. Lesson I.A.1
MANAGE: The function that prioritizes risks and allocates resources to treat, respond to, recover from and communicate about them. Lesson II.D.2
Manufacturing defect: A flaw where a particular unit departs from its intended design. Lesson II.B.5
MAP: The function that establishes context and identifies risks and impacts of an AI system. Lesson II.D.2
MAP function (NIST AI RMF): The NIST function that establishes context: intended purpose, settings, laws, norms and potential impacts. Lesson III.A.1
Market surveillance authority: A national authority that monitors AI systems on its market and can investigate, order corrective action or withdrawal, and impose fines. Lesson II.C.9
Marking grace period: Extra time, until 2 Dec 2026, for generative AI systems placed on the market before 2 Aug 2026 to add machine-readable marking. Lesson II.C.10
Meaningful human involvement: Review by a person with the authority, competence and information to change an automated outcome, as opposed to rubber-stamping. Lesson II.A.4
Meaningful human review: Review by a person with the competence, information and authority to change an AI-informed decision. Lesson I.C.5
MEASURE: The function that analyzes, assesses, benchmarks and monitors AI risks and trustworthiness. Lesson II.D.2
Measurement bias: Bias from using a feature or label that is a poor or uneven proxy for what you actually care about. Lesson III.B.4
Membership inference: Probing a model's outputs to determine whether a specific person's data was in its training set. Lesson III.C.6
Memorization: A model's tendency to retain and reproduce parts of its training data, which can expose personal data. Lesson II.A.2
Misalignment: When an AI system pursues a measurable proxy or purpose that diverges from what its owners actually intended. Lesson I.A.2
MITRE ATLAS: A public knowledge base of real-world adversary tactics and techniques used against AI systems. Lesson IV.C.3
Model card: A model card is a short, standardized document that describes a trained AI model. It states what the model is for and not for, how it was built and evaluated, and how well it performs across groups. It also lists known limits, helping deployers, auditors and users decide whether the model fits their use. Full explainer · Lesson I.C.2
Model change notification: A contract term requiring the vendor to give advance notice of material changes to the model, often paired with version pinning. Lesson IV.B.2
Model drift (model decay): The decline in a deployed model's performance over time, usually caused by data or concept drift. Lesson III.C.4
Model extraction: Reconstructing a working copy of a model by systematically querying it and learning from its outputs. Lesson III.C.6
Model owner: The named person accountable for a model's performance, documentation and risks throughout its life. Lesson IV.B.5
Model registry: The single record of every model version, with its hash, dataset versions, evaluation results, owner, approvals and status. Lesson III.C.10
MSIT: South Korea's Ministry of Science and ICT, which leads implementation and enforcement of the AI Basic Act. Lesson II.C.12
Multimodal model: A model that processes more than one type of data, such as text together with images, audio or video. Lesson IV.A.2

N

Near miss: An incident where harm was narrowly avoided; it is still documented and investigated. Lesson III.C.7
NIST AI 600-1 (Generative AI Profile): A July 2024 cross-sectoral profile naming 12 risks unique to or worsened by generative AI, with actions mapped to the AI RMF. Lesson II.D.3
NIST AI RMF: NIST's voluntary AI Risk Management Framework (AI 100-1, January 2023), organized around GOVERN, MAP, MEASURE and MANAGE. Lesson II.D.2
Non-AI alternative: A human or non-automated route offered instead of an AI process; it should be feasible, clearly explained, easy to choose and no worse in service. Lesson IV.C.6
Non-AI fallback: A manual or rules-based process that takes over an AI system's work when it is switched off or retired. Lesson IV.C.8
Non-consensual intimate imagery (NCII): Intimate images of a person shared without their consent, whether authentic or AI-generated. Lesson II.C.13
Non-obvious bias: Unfairness that standard technical and legal checks miss, such as culturally loaded language or subtle proxies. Lesson I.B.2
Notified body: An independent, accredited organization designated by a Member State to assess the conformity of certain high-risk systems. Lesson II.C.5

O

OECD AI Principles: The 2019 (updated 2024) OECD Council Recommendation setting five values-based principles for trustworthy AI and five recommendations for governments. Lesson II.D.1
On-premise deployment: Running an AI system on servers the organization owns and operates itself. Lesson IV.A.3
Opacity (black box): When an AI system's internal reasoning can't readily be inspected or explained, even by its builders. Lesson I.A.3
Open-weights model: A model whose trained parameters are published for download, usually under a license that may still carry conditions. Lesson IV.A.2
Operational controls: Mechanisms built into a running system, such as overrides, stop functions, fallbacks, escalation thresholds and logs, that let humans steer it. Lesson III.A.4
Operationalize: Turning a principle into concrete policies, procedures, controls and evidence. Lesson I.A.5
Optionality: Keeping feasible alternatives open, including non-AI options and opt-outs to a human route, and avoiding lock-in. Lesson III.A.8
Organizational harm: Damage the organization itself absorbs: disruption, breaches, fines, lost business, reputational or cultural damage, and acceleration risk from rushing AI out untested. Lesson I.A.2
Organizational readiness: Whether the organization has the leadership, budget, data governance, security and privacy maturity, risk tolerance and operating capacity to support an AI deployment. Lesson IV.A.1
Out-of-cycle review: An immediate policy review triggered by AI entering an existing program, a new AI program, or a new regulation. Lesson I.C.3
Out-of-scope use: A use the deployer or provider has identified as inappropriate or unsupported, documented so it can be prevented. Lesson IV.C.5
Overfitting: When a model memorizes its training data and performs poorly on new data. Lesson III.B.3
Override rate: How often human reviewers reject or change the system's output; a rising rate is an early warning sign. Lesson III.C.3

P

Performance requirements: The accuracy, speed, explainability, availability and error tolerance a specific use demands, set by the deployer for its own context. Lesson IV.A.1
Performance testing: Testing whether a system meets speed and scale targets, such as latency, throughput and resource use under load. Lesson III.A.7
Permissive license: An open license, such as MIT or Apache 2.0, that allows commercial use and modification as long as notices are kept. Lesson IV.B.4
Personal data breach: A security breach leading to accidental or unlawful destruction, loss, alteration, unauthorized disclosure of, or access to, personal data. Lesson II.A.4
Plan-Do-Check-Act (PDCA): The continual-improvement cycle behind ISO management-system standards. Lesson II.D.4
Planned retirement: Ending an AI system's use on a schedule for reasons such as obsolescence, vendor end of support, a replacement or a changed business need. Lesson IV.C.8
Positionality exercise: A structured reflection in which a team states how its members' backgrounds and roles shape what they notice, to surface blind spots before design choices. Lesson III.A.8
Post-deployment monitoring plan: A deployer's documented plan covering metrics and thresholds, user feedback, appeals and overrides, incident response, change management and decommissioning (NIST AI RMF MANAGE 4.1). Lesson IV.C.4
Post-market monitoring: A provider's ongoing, documented collection and analysis of a system's real-world performance throughout its life (Art. 72). Lesson II.C.5
Post-market monitoring plan: The documented plan for how a provider will collect and analyze real-world performance data; part of the technical documentation (Art. 72). Lesson III.C.9
Pre-deployment pilot: A limited, controlled trial with real users and pre-agreed success criteria before the go/no-go decision. Lesson III.A.6
Pre-mortem: A forecasting exercise that imagines a future failure and works backward to what caused it. Lesson IV.C.5
Precision and recall: Precision is the share of flagged cases that are truly positive; recall is the share of true positives the model catches. Lesson III.B.3
Preemption: Federal law overriding conflicting state law; it requires Congress, a valid federal regulation or a court, not an executive order alone. Lesson II.C.13
Principle trade-off: A situation where advancing one responsible-AI principle, such as accuracy, costs another, such as fairness or privacy. Lesson I.A.5
Privacy by Design (PbD): Ann Cavoukian's framework of seven principles for building privacy into systems proactively and by default. Lesson II.A.2
Probabilistic output: A likelihood or best guess that can vary and can be wrong, rather than a guaranteed result. Lesson I.A.3
Processor: A party that processes personal data on a controller's behalf and documented instructions. Lesson II.A.3
Product liability: Liability for harm caused by defective products, which under the revised EU Product Liability Directive extends to software, including AI. Lesson IV.B.3
Product Liability Directive (EU) 2024/2853: The EU's revised no-fault product liability law, which treats software and AI as products and eases the burden of proof. Lesson II.B.5
Profile (current vs target): A tailored application of the AI RMF; comparing a current profile with a target profile reveals gaps to address. Lesson II.D.3
Profiling: Automated processing of personal data to evaluate aspects of a person, such as work performance, finances, health, preferences or behavior. Lesson II.C.3
Prohibited AI practice: An AI use banned outright by Art. 5 because it's incompatible with EU values and fundamental rights. Lesson II.C.2
Prohibited inputs: Categories of data, such as confidential, personal or regulated information, that must never be entered into external AI tools. Lesson I.C.5
Prompt engineering: Shaping a model's behavior through the instructions and examples sent with each request, without changing the model. Lesson IV.A.4
Prompt injection: An attack that hides malicious instructions in a model's input to hijack its behavior. Lesson I.C.3
Proportionality: Scaling governance effort to the organization's context and the risk of each AI use. Lesson I.B.4
Proprietary (closed) model: A model whose code and weights are not published; you use it under license or through an API. Lesson IV.A.2
Provider: The party that makes an AI system available to others; under the EU AI Act, it places the system on the market or puts it into service under its own name. Lesson I.B.5
Proxy variable: A seemingly neutral input, such as ZIP code, that correlates closely with a protected characteristic. Lesson II.B.2
Pseudonymization: Replacing identifiers with codes kept separately; the data stays personal data for anyone who can re-link it. Lesson II.A.2
Purpose limitation: The principle that personal data collected for specified purposes must not be reused in ways incompatible with them. Lesson II.A.1
Putting into service: Supplying an AI system for first use to a deployer, or for the provider's own use, in the EU for its intended purpose. Lesson IV.B.3

Q

Quality management system (QMS): A provider's documented policies and procedures ensuring a high-risk system is designed, built, tested and monitored in compliance with the Act (Art. 17). Lesson II.C.5

R

RACI: A chart that marks who is Responsible, Accountable, Consulted and Informed for each decision or task. Lesson I.B.1
Readiness assessment: The formal check after development and before deployment that decides whether a system is ready to release (go/no-go). Lesson III.C.1
Readiness review: The go/no-go check before deployment that confirms testing, documentation, oversight, training and incident plans are in place. Lesson I.C.2
Real-time remote biometric identification: Identifying people at a distance, live, by comparing their biometric data (usually faces) with a database, without their active involvement. Lesson II.C.2
Reasonable accommodation: An ADA-required adjustment, such as an alternative assessment, that lets a qualified person with a disability compete fairly. Lesson II.B.2
Reasonably foreseeable misuse: Use of an AI system in a way its provider didn't intend but which could result from predictable human behavior or interaction with other systems. Lesson II.C.4
Record of processing activities (ROPA): The Article 30 register of an organization's processing purposes, data, recipients, transfers, retention and security. Lesson II.A.3
Red teaming: Deliberately attacking your own AI system, as an adversary would, to find weaknesses before release. Lesson I.C.2
Redress: The means for people to report, contest and get correction of an AI system's harmful or wrong outcomes, designed in from the start. Lesson III.A.8
Regression testing: Re-running tests after any change to data, code, features, hyperparameters or thresholds, to confirm earlier results still hold. Lesson III.B.8
Reproducibility: The ability for others to obtain consistent results by repeating the documented process. Lesson III.B.6
Requirements gathering: Collecting what a system must do and how it must behave, including legal, fairness, privacy and explainability needs. Lesson III.A.3
Reservation of rights (TDM opt-out): A rightsholder's machine-readable statement that its online works may not be mined under the EU's general TDM exception. Lesson II.B.1
Residual risk: The risk that remains after mitigations, which someone with the right authority must formally accept. Lesson I.C.1
Responsible AI license (RAIL): A license that permits use of a model or code but bans listed harmful uses and requires those restrictions to be passed downstream. Lesson IV.B.4
Retention schedule: A documented set of rules saying how long each type of record is kept and when it must be destroyed. Lesson IV.C.8
Retraining trigger: An automated threshold or alarm that starts the retraining process when monitored performance crosses a defined line. Lesson III.C.4
Retrieval-augmented generation (RAG): A technique where the system retrieves relevant documents at query time and adds them to the prompt, leaving the model's weights unchanged. Lesson IV.A.4
Revoked executive order: A presidential order cancelled by a later order, as EO 14110 was in January 2025; it no longer directs federal agencies. Lesson II.C.15
Right of access: The right to obtain one's personal data and information about its processing, including meaningful information about the logic of automated decisions. Lesson II.A.4
Right to explanation (Art. 86): An affected person's right to a clear, meaningful explanation from the deployer of the AI's role in a significant decision based on an Annex III high-risk system. Lesson II.C.6
Right to object: The GDPR right (Article 21) to object to processing based on legitimate interests or public task, which the controller must honor unless it has compelling overriding grounds. Lesson II.A.1
Risk management system (Art. 9): A continuous, iterative process across the life cycle to identify, evaluate and mitigate risks to health, safety and fundamental rights. Lesson II.C.4
Risk mitigation hierarchy: An ordered preference for responses: avoid the risk, then reduce it, then restore, and compensate as a last resort. Lesson III.A.5
Risk register: The living record of identified risks with their likelihood, severity, owner and mitigations. Lesson III.C.7
Risk tier: A category (such as prohibited, high, medium or low) that sets how much review and control an AI use receives. Lesson I.C.1
Risk tolerance: How much AI-related risk an organization is willing to accept in pursuit of its objectives. Lesson I.B.4
Risk transfer: Shifting a risk's financial consequences to another party, such as an insurer or contractor; it doesn't reduce the harm to affected people. Lesson III.A.7
Risk-based approach: Regulating AI in proportion to the harm a use could cause, so obligations rise with the risk to health, safety and fundamental rights. Lesson II.C.1
Risk-driven deactivation: Switching off or restricting a live AI system because of a regulatory, performance, safety or security problem, under the deactivation policy. Lesson IV.C.8
Robotic process automation (RPA): Software that mimics repetitive human steps using fixed rules. It automates tasks but does not learn or infer. Lesson I.A.1
Robustness: A system's ability to maintain performance across varied, noisy, unexpected or manipulated conditions. Lesson III.C.8
Rogue deployment: A model running in production that didn't come through the approved pipeline or isn't in the registry. Lesson III.C.10
Role-based training: Training whose depth and content match what a person's role requires, layered on a shared baseline. Lesson I.B.3
Rollback: Returning a system to its last known-good model version or configuration. Lesson IV.C.7
Rollback plan: A tested way to return quickly to the previous version if the new one misbehaves. Lesson III.C.1
Root cause analysis: A structured investigation that traces an incident past its symptoms to the underlying cause that must be fixed. Lesson III.C.8

S

Safety component: A part of a product or system that performs a safety function, or whose failure endangers health, safety or property. Lesson II.C.3
Sampling bias: Bias from a training sample that doesn't represent the population the system will serve. Lesson III.B.4
Scope creep: An AI system gradually being used for purposes beyond those it was assessed and approved for. Lesson I.C.2
Secondary use: Use of an AI system, or its data, for a purpose beyond the one that was assessed and approved. Lesson IV.C.5
Security (penetration) testing: Structured testing of technical controls, such as APIs, infrastructure and access, to find exploitable vulnerabilities. Lesson III.C.5
Segregation of duties: A control ensuring no single person can build, approve and deploy a change alone. Lesson III.C.10
Sensitive data (US state laws): Categories such as biometrics, health, precise location and children's data that state privacy laws usually allow only with opt-in consent. Lesson II.A.5
Serious incident: An incident or malfunction leading to death or serious harm to health, serious disruption of critical infrastructure, breach of fundamental-rights law, or serious harm to property or the environment. Lesson II.C.5
Shadow AI: Shadow AI is the use of AI tools for work that the organization has not approved, assessed or recorded. A common case is staff pasting client data into a personal chatbot account. It creates hidden privacy, security, intellectual property and compliance risks because nobody has vetted the tool or knows what data it receives. Full explainer · Lesson I.C.5
Shadow deployment: Running a new model on live traffic without acting on its outputs, to compare it safely against the current system. Lesson III.C.1
Shadow mode: Running an AI alongside the existing process so its outputs can be compared without driving real decisions. Lesson III.A.6
Shared responsibility model: The split of security and compliance duties between a cloud provider (infrastructure) and its customer (data, access and use). Lesson IV.A.3
Sign-off: A named, accountable person's recorded approval that release criteria are met and the residual risk is accepted. Lesson III.C.1
SME penalty rule: For SMEs and start-ups, each AI Act fine is capped at the lower of the fixed amount and the turnover percentage (Art. 99(6)). Lesson II.C.9
SOC 2 Type II report: An independent auditor's report on whether a service provider's controls operated effectively over a period, typically 6 to 12 months. Lesson IV.B.4
Social scoring: Evaluating people over time based on behavior or personal traits in a way that leads to unjustified or disproportionate detrimental treatment. Lesson II.C.2
Societal harm: Damage to shared goods like democratic participation, trust in information, access to education or the environment. Lesson I.A.2
Special categories of personal data: GDPR Article 9 data (ethnic origin, political opinions, beliefs, union membership, genetic, biometric for identification, health, sex life or orientation) whose processing is prohibited unless an exception applies. Lesson II.A.5
Stakeholder mapping: Identifying everyone affected by or influencing a system, including people who never use it. Lesson III.A.5
Stakeholder salience: How much weight a stakeholder's voice should get, judged by power, legitimacy, urgency and how heavily the system affects them. Lesson III.A.8
Standard contractual clauses (SCCs): European Commission-approved contract terms used as a safeguard for transferring personal data out of the EU. Lesson II.A.3
Statement of Applicability: A 42001 document listing which Annex A controls an organization applies and why any are excluded. Lesson II.D.4
Storage limitation: Keeping personal data in identifiable form no longer than the purpose requires. Lesson II.A.2
Strict liability: Liability for harm caused by a defective product without proof that the maker was negligent. Lesson II.B.5
Subprocessor: A third party the vendor engages to process the customer's data, such as a hosting provider. Lesson IV.B.2
Substantial modification: A significant change to an AI system after it reaches the market; under EU AI Act Article 25 it can turn a deployer into a provider. Lesson I.B.5
Substantiation: The reasonable basis, such as test evidence, a company must have before making objective product claims. Lesson II.B.4
Synthetic data: Artificially generated records that mimic real data, used to fill gaps, cover rare cases and improve accuracy. Lesson III.B.5
Systemic bias: Bias embedded in institutions and society that flows into data even when each step looks neutral. Lesson III.B.4
Systemic risk: Risk from high-impact GPAI capabilities that could significantly affect the EU market, public health, safety, security or fundamental rights at scale; presumed above 10²⁵ FLOPs. Lesson II.C.8

T

TAKE IT DOWN Act: A 2025 federal law criminalizing publication of non-consensual intimate imagery, including AI deepfakes, and requiring platforms to remove it within 48 hours of a valid request. Lesson II.C.13
Technical documentation (Annex IV): The dossier a provider draws up before market placement to show a high-risk system meets the requirements. Lesson II.C.4
Temporal bias: Bias from data that is unbalanced across time, so it no longer reflects current reality. Lesson III.B.4
Test plan: A document written before testing that sets objectives, scope, scenarios, metrics, pass/fail criteria and roles. Lesson III.B.8
Test report: A dated record of what was tested, how, against which criteria, the results including failures, remediation and sign-off. Lesson III.B.6
Test set: Data kept completely separate and used once, at the end, to estimate real-world performance. Lesson III.B.3
TEVV: Test, evaluation, verification and validation: NIST's umbrella term for the checks run across the AI life cycle. Lesson III.B.7
Text and data mining (TDM): Automated analysis of text and data to find patterns; the EU allows it under two copyright exceptions. Lesson II.B.1
Third-party AI risk: Risk from AI systems, models, data or components supplied by vendors and their own suppliers. Lesson I.C.4
Threat hunting: Proactively searching a live system for signs that someone is already abusing or compromising it, rather than waiting for alerts. Lesson III.C.5
Threat modeling: Structured analysis, on paper, of a system's assets, entry points, likely attackers and techniques, used to prioritize defenses and scope testing. Lesson III.C.5
Three lines model: A governance pattern splitting risk ownership (first line), oversight (second line) and independent assurance (third line, internal audit). Lesson I.B.1
Tiered review: Scaling the depth of governance review to a use case's risk, so low-risk uses move fast and high-risk ones get full scrutiny. Lesson I.B.2
Trade secret: Confidential business information, such as model weights or curated data, protected for as long as reasonable steps keep it secret. Lesson IV.B.5
Training-content summary: A public, sufficiently detailed summary of the data used to train a GPAI model, prepared using the Commission's template (Art. 53(1)(d)). Lesson II.C.8
Transfer impact assessment: An analysis of whether the destination country's laws undermine the protection a transfer mechanism promises. Lesson II.A.3
Transparency: Disclosing what an AI system is, what it does and that AI is in use; NIST frames it as answering 'what happened'. Lesson I.A.4
Transparency report: A public summary of a system's capabilities, limitations, testing and risk management, voluntary or required by some laws. Lesson III.C.9

U

Underfitting: When a model is too simple or undertrained to capture the pattern, performing poorly everywhere. Lesson III.B.3
Unfair discrimination (insurance): The state-law standard barring insurers from distinctions not justified by risk, or based on protected traits. Lesson II.B.3
Unfair practice: A practice causing substantial injury that consumers can't reasonably avoid and that isn't outweighed by benefits. Lesson II.B.4
Untargeted scraping: Indiscriminately harvesting facial images from the internet or CCTV to build or expand a facial-recognition database. Lesson II.C.2
Use case: A specific description of what an AI system will do, for whom, in what setting, and within what boundaries. Lesson III.A.1
Use-case assessment: The intake review that captures an AI proposal's purpose, owner, data and affected people, and triages it by risk. Lesson I.C.1
Use-case context: The business objectives, performance requirements, data availability, ethical considerations and workforce readiness that decide whether an AI deployment makes sense. Lesson IV.A.1
Use-case evaluation: Assessing whether a specific use of AI is lawful, beneficial and within risk tolerance before proceeding, and again whenever the use changes. Lesson III.A.6
User: The individual who interacts with an AI system or acts on its outputs. Lesson I.B.5

V

Validation set: Data held back from training and used during development to tune settings and compare model versions. Lesson III.B.3
Vendor due diligence: Risk-scaled investigation of a supplier's AI: training data, testing, documentation, security, data use and update practices. Lesson I.C.4
Verification vs. validation: Verification confirms a system meets its specification (built it right); validation confirms it fulfills the requirements for its intended use (built the right thing). Lesson III.B.7
Version pinning: Staying on a specific, validated model version until a newer one has been tested and approved. Lesson IV.C.2

W

Withdrawn proposal: A legislative proposal the European Commission formally abandons, such as the AI Liability Directive in 2025; it never becomes law. Lesson II.C.15
Workforce readiness: Whether the people who will use, oversee or be affected by an AI system have the skills, training, roles and trust to work with it well. Lesson IV.A.1