AIGP glossary

Reviewed by Victor Humenhuk. Published 31 August 2026. Reviewed 31 August 2026.

Review the recurring terms in this AIGP guide. Use each definition as a prompt, then explain the term without looking.

This glossary supports recall. The linked lessons provide the legal context, exceptions and scenarios that a short definition cannot hold.

21st Century Cures Act

Not AI-specific; promotes advanced tech plus accessibility and transparency of health data.

3×3 harms matrix

Rates each risk's severity and probability, multiplying them for a score.

A COD SHiP

Mnemonic for the 7 governance-critical AI characteristics: Autonomy, Complexity, Opacity, Data dependency, Speed and scale, Harm/misuse, Probabilistic outputs.

Acceleration risk

Unanticipated organisational harm from rapid AI processing and complexity, with wider/more severe impacts.

Acceptable use policy

Vendor rules on permitted uses, reviewed when screening third-party AI.

Accountability

Identified people/orgs answerable for an AI system's functioning and impacts; anchor of governance.

Accuracy

Share of model outputs that are correct against ground truth.

Active learning

Model flags the most informative unlabelled examples for a human to label.

Adaptive learning

Systems adjusting behaviour/content in response to new data and user interactions over time.

Adversarial attack

Deliberately crafted inputs designed to fool a model into errors or unsafe behaviour.

Adverse action notice

Notice in finance when an application is declined, implicating any AI in that process.

Adverse Impact Ratio (AIR)

A measure used to assess AI outputs for bias.

Agentic AI

AI agents that plan and execute multi-step tasks with significant autonomy, chaining tools.

AGI (Artificial General Intelligence)

Strong AI with human-level intelligence across many domains; does not yet exist.

AI agent

Goal-driven, autonomous performance of specific tasks with simple workflows (e.g. antivirus, robotic vacuums).

AI assurance

Measurable mechanisms demonstrating a system is trustworthy and compliant (evals, audits, certs, docs).

AI audit

Independent, systematic review of an AI system or its governance against defined criteria.

AI Basic Act

South Korea's Act on the Development of AI and Establishment of Trust; second national comprehensive AI law, effective Jan 2026.

AI Decommission Checklist

Governance checklist covering data, models, infrastructure, documentation, risk sign-off and post-mortem reviews.

AI governance

Rules, policies, processes and accountability structures steering responsible AI across its life cycle.

AI impact assessment (AIIA)

The severity lens, gauging how bad mapped risks are and potentially guiding the go/no-go decision.

AI Incident Database

A database of known AI incidents reviewed to grasp the breadth of potential problems.

AI Incident Database (AIID)

Database the CSET taxonomy characterises, recording harms, entities and technologies in AI incidents.

AI inventory

A central inventory of AI applications plus a repository of algorithms for visibility and transparency.

AI literacy

The skills and knowledge to engage with AI in an informed, responsible and effective manner; mandated by EU AI Act Article 4.

AI model (OECD dimension)

Computational representation of the environment: technical type, how it is built and used.

AI Omnibus

An amending regulation that entered into force on 27 July 2026 and changed parts of the AI Act, including the application dates for high-risk systems.

AI Promotion Act

Japan's innovation-first soft-law framework; government can investigate and advise but imposes no penalties.

AI registrar

A central record where incident and issue information is kept.

AI system development life cycle

Seven iterative stages from plan/design through to lawful decommissioning, with governance at each stage.

AI Verify

Singapore's tool validating AI systems against 11 ethics principles.

Algorithm

A defined sequence of computational steps or rules transforming inputs into outputs.

Algorithmic Impact Assessment (AIA)

Assessment covering data issues and documenting the stakeholder group's decisions and who accepts risk; Canada publishes an AIA tool.

America's AI Action Plan

July 2025 US plan with 90+ federal actions, three new EOs and three pillars (innovation, infrastructure, leadership).

ANI (Artificial Narrow Intelligence)

Narrow/weak AI that excels at specific tasks but cannot transfer knowledge; exists everywhere today.

Anonymisation

Removing items that could identify individuals, such as names and addresses; complete anonymisation is difficult.

Appropriation

Reusing data consented for one purpose to train AI; an ethics-of-consent problem.

Article 22

GDPR provision: a general prohibition (with three exceptions) on decisions based solely on automated processing with legal/significant effects.

Article 35 (DPIA)

GDPR provision mandating data protection impact assessments for high-risk or significant processing.

Artificial general intelligence (AGI)

Hypothetical AI matching or exceeding human cognitive ability across virtually any task.

Artificial intelligence

Machine systems that infer from inputs how to generate outputs influencing real or virtual environments.

Artificial intelligence (AI)

A field of computer science that uses computational techniques to simulate intelligent behaviour and automate tasks.

ASI (Artificial Super Intelligence)

Hypothetical AI surpassing humans in virtually every domain; feasible only if AGI is achieved first.

Association rule learning

Unsupervised technique identifying relationships between data points (people who buy X also buy Y).

Automated decision-making (ADM)

Decisions by technological means without meaningful human involvement; GDPR Article 22.

Automation bias

Over-trusting machine decisions and assuming they are always correct.

Autonomy

The degree to which a system operates and decides without human intervention.

Bartz v Anthropic PBC (2025)

Significant US ruling on whether developers can train AI models on copyright-protected works.

Bayesian Improved Surname Geocoding

The prominent proxy method to infer demographics from less sensitive data (surname plus geography).

Benchmarking

Standardised tests comparing systems on accuracy, speed and complex-task handling; useful for black-box models, executed once a candidate model exists.

Bias

Systematic error producing unfair outcomes; computational, cognitive and societal flavours.

Bias-testing dilemma

Minimisation discourages holding sensitive data, but bias testing needs it - creating pressure to collect it for evaluation.

Blueprint for an AI Bill of Rights

2022 US guidance setting five principles for automated systems plus concrete steps.

Bootstrap aggregating (bagging)

Training models on random resamples (with replacement) and combining outputs to reduce variance.

Breach of warranty

US product-liability claim for failure to fulfil promises made about a product.

Brittleness

An AI performing successfully in one instance yet failing in another.

Broad AI

Intermediate step beyond ANI, often a group of AI systems working together, e.g. AI agents, autonomous vehicles.

Bug bounty

Programme rewarding external discovery of flaws, used for monitoring engagement and feedback.

Business operator

South Korea's role term covering AI Development and AI Utilization operators (replaces provider/deployer).

C-L-EC

Article 22 exceptions: necessary for a Contract, authorised by Law, or based on Explicit Consent.

CAC

Cyberspace Administration of China; oversees AI services, requiring filings, security reviews and algorithm registration.

California AB 2013

Law requiring GenAI training-data transparency (1 Jan 2026).

California BOT Act

Prohibits undisclosed bots that encourage a sale.

California SB 942

AI Transparency Act on watermarking and detection of AI content (1 Jan 2026).

California TFAIA (SB 53)

Frontier AI Act regulating the most advanced systems for transparency, safety, accountability and responsiveness (1 Jan 2026).

Category 1 adoption

An existing function performed a new way via AI - prior regulatory requirements continue to apply.

Category 2 adoption

A new process made possible by AI - the question is how existing requirements apply to it.

CE marking

EU conformity mark importers must confirm on high-risk AI alongside the conformity assessment and database registration.

Centralised governance

One team or person responsible for AI affairs, with everyone flowing through that point.

CFPB

US bureau requiring creditors to explain specific reasons behind adverse credit decisions, even from 'black box' models.

Challenger model

A new model tested against the production 'champion' on the same data.

Champion model

The production-proven model a challenger is tested against.

Champion vs challenger

Developing a challenger model to test against the existing champion model to assess drift and unexpected results.

Chatbot

Software simulating human conversation via text or voice; now commonly LLM-powered.

Citron & Solove taxonomy

Seven privacy harm types: physical, reputational, relationship, economic, discrimination, psychological, autonomy.

Classification model

Predicts which discrete category an input belongs to.

Classification models

Supervised models predicting categorical responses, e.g. spam vs not spam (SVM).

Cloud-based deployment

Hosting where a third-party provider handles the infrastructure; easy to scale but adds latency and third-party data risk.

Clustering

Unsupervised grouping of data points by similarity with no predefined labels.

CNN

Convolutional neural network specialised for images.

Collection limitation

A subset of minimisation restricting how much and what kind of data is gathered.

Colorado AI Act (SB 24-205)

Law on consequential decisions and AIAs (1 Feb 2026), facing a proposed overhaul.

Compounded systemic impact

When a single agent error spreads quickly across departments and platforms.

Compute

Processing power and hardware (GPUs) to train/run AI; a cost driver and frontier-model policy lever.

Computer vision

Field enabling machines to derive meaning from images and video.

Conformity assessment

Process showing a system meets legal/standards requirements; EU AI Act gate for high-risk before market.

Conformity assessment (CA)

Pre-market evaluation demonstrating a high-risk AI's compliance; underpinned by technical documentation.

Consequential decisions

Colorado SB 24-205 term for high-risk decisions on jobs, loans, housing and care, triggering an AIA.

Containerisation

Packaging a model and all its dependencies into a self-contained unit to ease deployment and reduce compatibility issues.

Contestability

People's ability to challenge an AI-influenced decision and obtain review or redress.

Corpus

Large, structured collection of text or speech used to train language models.

Counterfactual

Explanation showing the minimal input change that would flip the output.

Counterfactual explanation

Detail of what new or different input would change the output of the AI process.

CSET AI Harm Taxonomy

Georgetown CSET framework for the AI Incident Database; defines AI harm with four elements, all four must be present.

Data

Raw information used to train AI models: text, images, audio, video, sensor data.

Data and input

OECD dimension covering data and expert input, collection method, structure, training and production data.

Data cleansing

Removing erroneous and irrelevant data to protect performance and reduce privacy risk.

Data controller

The party that decides what and how personal data is processed; the GDPR applies whether a human or AI processes it.

Data drift

Input data's statistical properties change over time vs training data; performance degrades.

Data labelling

Tagging or annotating data (images, audio, text); accuracy directly affects learning.

Data leak

Out-of-scope data leaking in to inflate performance, or a model exposing sensitive training data.

Data life cycle

Collection, Use, Disclosure, Retention, Destruction - the stages personal data passes through.

Data lineage

Traceability of the origin and nature of training data, often ambiguous with third-party models.

Data localisation laws

Jurisdictional requirements governing where data may be physically stored.

Data minimisation

Data must be adequate, relevant and limited to what is necessary; avoid 'nice to have' data.

Data persistence

Privacy risk where data outlives its original purpose.

Data poisoning

Corrupting the training data so the model learns wrong or malicious behaviour.

Data provenance

Documented history and origin of data; supports integrity and identifies applicable laws.

Data quality

How accurate, complete, relevant, representative and fit-for-purpose the data is.

Data spillover

Privacy risk where unintended individuals' data gets collected.

Decentralised governance

'Local governance', decision-making delegated to lower levels with a bottom-to-top flow and wider span of control.

Decision tree

Flowchart-like model splitting data on feature questions until a leaf prediction; interpretable.

Decision trees

Supervised learning algorithm for classification and regression.

Decommissioning

Retiring an AI system when no longer needed, valuable or current; sensitive data archived or destroyed lawfully.

Deep learning

ML using many-layered neural networks; the engine behind GenAI.

Deep learning (DL)

ML using multi-layered neural networks; processes unstructured data and reduces feature engineering.

Deep Synthesis Provisions

China's rules requiring labelling and watermarking of deep-synthesis (deepfake) AI outputs.

Deepfake

AI-generated impersonation used for fraud, blackmail or misinformation.

Deepfakes

AI-generated/manipulated audio-visual content realistically depicting fabricated acts or speech.

Deployer

Entity or professional user applying an AI system for a specific purpose; responsible for safe, ethical use and monitoring.

Deployment

The transition of an AI system from a development and testing environment to a real-world, operational setting.

Destruction

Making personal data unrecoverable at the end of the data life cycle.

Deterministic

Software behaviour where the same input always yields the same output.

Developer

Role that designs, builds and tests models and hands deployers documentation on uses, limits and training data; called 'developer' in the Colorado AI Act.

Differential privacy

PET that blurs information within datasets so data stays meaningful but individuals can't be identified.

Diffusion model

GenAI that reverses gradually added noise to generate images or audio.

Digital India Act

Proposed Indian law replacing the IT Act 2000 to address high-risk AI, complementing the DPDP Act.

Digital Services Act

EU law requiring recommender-system and ad-targeting transparency, overlapping the GDPR.

Discriminative model

Learns the boundary between classes to classify; distinguishes rather than creates.

Disinformation

False content spread deliberately, with intent to deceive.

Distributor

Any entity other than the provider or importer that makes AI available on the market through the supply chain.

Distrustful AI

The mirror of trustworthy AI: black-box decisions, unfair outcomes, no explainability, diminished human experience.

Domestic representative / agent

Local appointee required of foreign AI operators above thresholds (South Korea, EU representative equivalents).

DPIA

Data Protection Impact Assessment - identifies and minimises risks from processing personal data.

Economic context

OECD dimension covering sector, business model, critical/non-critical nature, deployment, scale and maturity.

Edge cases & outliers

Rare data points outside the normal range causing mistakes the model was never trained for.

Edge computing

Processing data near its source; pushed by autonomous vehicles' need for fast perception and decisions.

Edge deployment

Hosting on edge devices like smartphones; low latency and privacy but limited compute.

EDPB opinion (2024)

European Data Protection Board opinion (prompted by the Irish DPA) on anonymity, legitimate interest and unlawful training data for AI models.

EEOC Guidance on AI and Hiring (2021)

AI must comply with federal nondiscrimination laws and not disproportionately disadvantage protected groups.

Effective challenge principle

Experts get the chance to challenge the risk model to expose limitations and improve it.

Entropy

Measure of uncertainty/randomness in data or predictions; used for decision-tree splits.

EO 14179

US 'Removing Barriers' executive order (23 Jan 2025) replacing the rescinded EO 14110.

Ethics by design

Sibling of privacy by design: ethical issues resolved at the start and reassessed as risks evolve; must be continuous.

EU AI Act

The 2024 landmark, risk-based, extraterritorial regulation governing how AI is used, not the technology itself.

EU AI Act Article 4

Provision mandating providers and deployers ensure a sufficient level of AI literacy for relevant staff.

EU AI Board

EU advisory body providing technical guidance on AI governance.

EU AI Office

The EU's central authority for supervising AI under the EU AI Act.

EU risk pyramid

Four tiers, unacceptable (banned), high (mandatory requirements), limited (transparency), minimal (voluntary).

EU two-way information chain

Providers brief deployers; deployers disclose to users and feed information back to the provider.

Event detection

AI use-case bucket spotting patterns that signal something happened, e.g. fraud, cyber events.

Executive champion

A leader appointed to drive support and alignment for the AI governance effort.

Existing laws principle

All existing sector and jurisdiction laws still apply when AI is used; AI usually adds new obligations on top.

Expert system

Early AI using a knowledge base plus an if-then inference engine to mimic specialists.

Explainability

Ability to describe in understandable terms why a model produced a particular output.

Explicit consent

Freely given, specific, informed and unambiguous agreement, with a means to opt out.

Exploratory data analysis (EDA)

Initial investigation of a dataset to spot patterns, anomalies and gaps before modelling.

Exposing the model

Making a model accessible to other systems/applications, commonly via REST APIs or embedding.

Fail-safe plans

Predefined mechanisms moving a failing system to a safe state: shutdown, fallback, human takeover.

Fair Housing Act (1968)

Foundational US housing nondiscrimination law; compliance required even when AI ranks or scores people.

Fair use

US doctrine permitting unlicensed use in cases like criticism, news reporting and research, judged case-by-case.

Fairness

Outputs free from unjustified differential treatment or impact across individuals and groups.

Fault liability

Regime where victims must prove an action/inaction (e.g. negligence) by the maker caused the harm.

Faulty inference

Unreliable AI accuracy causing data to be attributed to the wrong individual.

FCRA

Fair Credit Reporting Act - foundational US credit/lending law with no AI-specific guidance.

Feature

A specific measurable aspect or characteristic, e.g. height, colour or substance.

Feature engineering

Deciding which features matter - to improve performance, cut cost and boost explainability.

Feature flags

Mechanism to disable functionality without redeploying code and control availability to specific users/groups.

Federated learning

Local systems train a shared model on their own data; only results aggregate centrally.

Fictional design ownership scenario

AI-generated design ownership is ambiguous until governance and vendor terms define ownership and IP rights.

Fine-tuning

Further training a pre-trained model on a smaller, task/domain-specific dataset.

FIPs

Fair Information Practices, an example of foundational privacy/AI principles.

FIPs (Fair Information Practices)

Eight 1980 OECD privacy principles covering data collection, use, protection and individual rights; ancestor of AI governance principles.

Five harm targets

Individuals, groups, society, organisations and ecosystems; one AI system can harm several at once.

Five incident causes

Brittleness, lack of robustness, lack of quality data, insufficient testing, and model or data drift.

Five V's

Volume, Velocity, Variety, Veracity, Value - checks in data preparation/wrangling.

Forecasting

AI use-case bucket predicting future values, e.g. demand, surge pricing, weather.

Foundation model

Large model trained on broad data at scale, adaptable to many downstream tasks.

Frameworks

A means to operationalise principles, context-sensitive and never one-size-fits-all (e.g. ISO 42001, NIST AI RMF, HUDERIA).

Frontier models

The most advanced capabilities of AI, warranting heightened legal and risk attention.

FTC authority

Broad authority over 'unfair or deceptive' practices that already covers algorithms and AI.

FTC Section 5

US authority over unfair and deceptive practices, applied to AI/ML systems.

Fundamental Rights Impact Assessment (FRIA)

EU assessment of an AI's impact on fundamental rights, required of public bodies and public-service providers.

GDPR

The EU General Data Protection Regulation, in effect since 2018; technology-agnostic global baseline for data protection.

General-Purpose AI (GPAI)

Models trained for broad tasks that adapt into many downstream systems (LLMs, multimodal, vision).

Generalization

A model's ability to perform well on new, unseen data.

Generative AI

Models that create new content from patterns learned in training data.

Generative AI (GenAI)

DL models that generate new content representative of training data but distinctly unique.

Generative AI Profile

NIST companion document applying the RMF to generative AI.

GNN

Graph neural network specialised for graph data.

Goal misalignment

When an agent achieves its goal but not in the way creators intended.

Goal-driven optimisation

AI use-case bucket finding the best solution to a defined problem, e.g. supply chain, routing.

Greedy algorithms

Make the locally optimal choice each step; fast but not guaranteed globally optimal.

Ground truth

Verified real-world correct answer used to check labels and score accuracy.

Hallucination

An AI's wrong output that misidentifies or misses real information/threats.

Hallucinations

GenAI output contradicting the source or factually wrong while presented as fact.

Harms taxonomy

A list of negative consequences from data leak/misuse; an ontological map breaking harms into constituent components like attacker capacity and opportunity.

HAT test

Trustworthy AI is Human-centric, Accountable, Transparent; operates in an expected, legal and fair manner.

High risk / high-impact AI

AI significantly affecting rights, safety or access to essential services; allowed but under strict obligations.

Homomorphic encryption

A PET that can protect training and testing data.

HUD 2020 guidance

Automated rental and mortgage lending decisions must adhere to FHA nondiscrimination.

HUDERIA

Council of Europe's Human Rights, Democracy and the Rule of Law Impact Assessment for AI, risk-based and proportionality-driven.

Human disempowerment

Erosion of human expertise (skill atrophy) and dignity from over-reliance on agents.

Human in the loop (HITL)

Design keeping a human in the decision path to review, confirm or override outputs.

Human-centric AI

AI designed around human needs, values, dignity and wellbeing; augments rather than replaces.

Hybrid governance

A central entity holds main responsibility while local entities fulfil and support its policies; typical in large organisations.

IEEE 7000-2021

A framework for addressing ethical concerns during system design.

Illinois BIPA

Biometric Information Privacy Act imposing disclosure and oversight duties on biometric applications.

Impact assessment

Risk management tool assessing a system's benefits, risks and limitations across its life cycle.

Implicit bias

Unconscious discrimination stemming from biased data or developer assumptions.

Implied consent

Consent inferred from actions or context (e.g. entering a camera-monitored store) - weaker than explicit.

Importer

Entity bringing an AI system into the domestic market from a third country; under the EU Act must be established in the EU.

Indemnity problem

Standard IP indemnities exclude modifications, combinations and out-of-scope use - all inherent to AI, so they break down.

Individual participation

FIP: appropriate access so a person can obtain, amend, correct or challenge their data.

Inference

Production phase where a trained model applies what it learned to new inputs.

Inference engine

Expert-system component applying knowledge via a rule-based approach, often showing its reasoning.

Input data

Data fed into the system at use time, on which it bases output.

Intellectual property

Creations of the human mind used in commerce, protected by patents, copyright and trademarks.

Interim Measures for GenAI Services (2023)

China's rules for public generative AI; exempt research institutions, require security reviews and registration.

Interpretability

Degree a human can consistently understand and predict how the model works internally.

Interruptibility

The agentic safety property of being able to gracefully shut an agent down.

ISO 42001

AI management system standard with which decommissioning documentation should align.

ISO/IEC 22989:2022

Standard establishing AI concepts and terminology, defining over 100 key concepts for a shared vocabulary.

ISO/IEC 42001

AI management system standard underpinning foundational guardrails.

ISO/IEC 42001:2023

AI management system standard for using AI responsibly across any size and industry, integrating risk assessment/treatment and third-party management.

ISO/IEC 42005:2025

Standard giving structured guidance for conducting AI system impact assessments across the life cycle.

Knowledge base

Organised collection of facts from human experts in one domain within an expert system.

KYC

Know Your Customer - process by which financial institutions verify customers and check funding sources are legitimate.

Large language model (LLM)

Transformer-based foundation model trained on massive text to understand and generate language.

Large language models

LMs with billions to trillions of parameters (GPT-4), complex generalists.

Legal risk

AI risk from a complex web of laws, noncompliance, liability for harm, IP disputes, human rights violations and reputational damage.

Legitimate interest test

Three-step test - Interest: Necessity: Balance - to rely on legitimate interest as a lawful basis.

LGPD

Brazil's General Data Protection Law, which also protects sensitive/special categories of data.

Limited / transparency risk

Lower-risk AI subject to disclosure or labelling duties only (e.g. chatbots, watermarked GenAI outputs).

Linear regression

Algorithm making numeric predictions from continuous variables.

Logistic regression

Probabilistic algorithm giving the likelihood of a usually binary outcome.

Machine learning

AI subset where systems learn patterns from data without explicit rule-by-rule programming.

Machine learning (ML)

Algorithms that learn patterns from data and improve over time without explicit programming.

Machine learning model

Trained artifact of learned parameters/logic used to predict on new inputs.

MAESTRO

A multi-agent framework cited for modelling emergent agentic behaviours.

Manage (NIST)

Phase that prioritises and acts (mitigate, transfer, avoid or accept) and monitors.

Map (NIST)

Phase that establishes context and identifies risks within it.

Measure (NIST)

Phase that assesses, analyses and tracks mapped risks with metrics.

Minimal / no risk

Low-concern AI (spellcheck, spam filters, games) subject to voluntary standards and codes at most.

Misinformation

False content spread without intent to deceive; wrong, not malicious.

MITRE PANOPTIC

Privacy taxonomy combining contextual domains and privacy activities; data-driven, supports threat assessment, risk modelling and red teaming.

Model

A program that applies algorithms to data to predict or decide based on learned patterns.

Model AI Governance Framework

Singapore's 2019 voluntary framework (Asia's first) on explainable, fair and human-centric AI.

Model card

Standardised documentation of one model: purpose, data, versions, metrics, bias/explainability reports.

Model Card Regulatory Check app

Tool that automates compliance checks from model cards.

Model cards / fact sheets

Standardised information on the model, its function and output, including version and dataset used.

Model drift

When the relationship between input data and output predictions changes over time, degrading performance.

Model inversion / extraction / poisoning / evasion

Four AI-specific security threats existing protocols miss.

Multimodal models

Process and/or generate more than one data type (text, images, audio, video).

Multimodal models (LMMs)

Models handling inputs/outputs across image, video, audio and text, with NLP as a key component.

NAIC Model Law (2020)

Insurance model law: algorithms must not perpetuate unlawful or unethical discrimination.

National AI Committee

South Korean body that recommends to agency heads and decides major AI policies.

Natural language processing (NLP)

Field enabling machines to understand, interpret and generate human language.

Neural networks

Layered interconnected nodes passing weighted signals to learn complex patterns.

NIST AI RMF

The NIST AI Risk Management Framework underpinning foundational guardrails.

NIST ARIA

Assessing Risks and Impacts of AI, a system assessing LLMs against scenarios to confirm capabilities, red-team guardrails and field-test use.

NIST Core functions

Govern, map, measure, manage, the four functions of the NIST AI RMF Core.

NIST Playbook

Companion suggesting actions to accomplish the outcomes of the Core functions.

NYC Local Law 144

Mandates AI bias audits for automated employment decision tools.

OECD AI Principles

An intergovernmental set of AI values, an example of principles.

OECD Framework for the Classification of AI Systems

A user-friendly framework that classifies AI systems and examines their risks across five dimensions.

Office for Responsible AI

A dedicated AI governance body/function within an organisation's stakeholder structure.

On-premise deployment

Hosting on the organisation's own servers and hardware; greater control but greater upfront cost.

Open source models

Models publicly available to use, modify and distribute; risk lies in quality control and security.

Open-source software

Software with publicly available source code to use, modify and distribute under licence.

Operational risk

AI risk of high costs (processors, large datasets, fast networks), skilled hires, environmental footprint and data corruption/poisoning.

Overfitting

Model memorises training data (incl. noise); great in training, poor on new data.

Oversight

Supervision arrangements, human or institutional, that monitor, review and can intervene.

Oversight body

A cross-functional, demographically diverse body reviewing higher-risk AI use cases in ethical grey areas.

Parameters

Internal values a model learns during training; billions for LLMs, a proxy for scale.

PEDMT

Mnemonic for the 5 OECD dimensions: People & planet, Economic context, Data & input, AI Model, Tasks & output.

People and planet

OECD dimension covering who is affected (human rights, environment, society); where privacy sits.

PETs

Privacy-enhancing technologies; digital solutions letting data be used while protecting confidentiality, preventing intentional and accidental misuse.

PETs (privacy-enhancing technologies)

Technologies addressing data security and privacy concerns; AI drives the need for them.

PIA

Privacy Impact Assessment - analyses how personally identifiable information is handled and confirms privacy compliance.

Post processing

Adjusting model outputs after inference (thresholding, filtering, bias correction).

Pre-deployment pilot

A late-stage validation activity under realistic, production-matching conditions immediately before go-live.

Preprocessing

Cleaning and transforming raw data before training (dedup, normalise, encode, fill gaps).

Principles

A set of values/guidelines enabling consistency and responsible AI use, similar around the world (e.g. OECD, UNESCO, FIPs).

Privacy by design

Apply data protection from the initial planning stage and process by default only the data necessary for each purpose.

Privacy risk

AI risk from data persistence and data spillover, complicated consent; countered by data minimisation, transparency and GDPR.

Probabilistic outputs

Likelihood-based results that can vary for the same input and carry uncertainty.

Probability/severity harms matrix

Risk strategy rating severity and probability and multiplying them for a risk score.

Prohibited risk

AI uses banned outright as threats to rights or safety (e.g. social scoring, manipulation, real-time public facial recognition).

Prompt

The input or instruction given to a GenAI system to elicit an output.

Prompt engineering

Crafting and refining prompts to steer model output quality, format and behaviour.

Proportionality principle

HUDERIA's methodology principle matching the assessment effort to the level of risk.

Proprietary model

A model an organisation both develops and deploys, creating a dual provider/user role and liability.

Proprietary models

Models developed by specific organisations with restricted access; can limit transparency and auditing.

Provider

Entity that develops and makes an AI system available on the market; carries the most extensive, whole-life-cycle obligations.

Pseudonymisation

Replacing identifiers so data is still personal information and GDPR obligations still apply; drops AI data utility.

Purpose limitation

Collect and use personal data only for the specified, determined, legitimate and clear purpose.

Purpose specification

FIP: disclose specific purposes up front, then use data only for compatible purposes.

Purpose specification & minimisation

Keeping data unnecessary for the application out of model training to protect privacy.

RAG

Retrieval-augmented generation: pulls in external information to boost LLM accuracy and relevance.

Random forest

Ensemble of many decision trees voting/averaging to boost accuracy and tame overfitting.

Random forests

Ensemble of decision trees, more accurate and better at complex data than a single tree.

Rebuttable presumption of defectiveness

PLD device letting courts infer a defect/causality and shift the burden to the manufacturer.

Recital 26

GDPR recital emphasising pseudonymisation and anonymisation to safeguard personal data.

Recognition

AI use-case bucket identifying images, faces, speech and text patterns.

Record of processing activities

A controller record kept with at least the regulatory minimum information on how the system operates.

Red teaming

Simulating adversarial attacks to expose flaws, bias and misinformation before release.

Red-teaming

Adversarial testing required of systemic-risk GPAI to probe for vulnerabilities and harmful behaviour.

Regression models

Supervised models predicting continuous numerical outcomes, e.g. car price (SVR).

Regulation by guidance

Japan's tradition of nonbinding, cooperative oversight over punitive enforcement.

Reidentification

Recombining personal data from multiple sources to compromise privacy.

Reinforcement learning

An agent learns by trial and error, optimising for rewards and penalties.

Reinforcement learning with human feedback (RLHF)

Uses human preference ratings as the reward signal to align outputs.

Release readiness assessment

Go/no-go check that the system works, tested well, conforms, has clean data and a model card.

Reliability

Consistent, correct performance over time and across real-world conditions.

Retrieval augmented generation (RAG)

Grounds LLM output by retrieving from a knowledge base beyond the training data.

Retrieval-augmented generation (RAG)

Optimising LLM output by referencing a knowledge base beyond the training data.

Revised Product Liability Directive

Directive 2024/2853, in force by December 2026; eases proof and compensation for AI-caused harm.

Risk formula

Probability of the harm × potential (severity) of it happening.

Risk mitigation hierarchy

The 'now what' used with the matrix - avoid, minimise, remediate and/or offset a risk's impact.

Risk-based classification

Sorting AI into tiers (prohibited, high, limited, minimal) and scaling obligations to the level of risk.

RNN

Recurrent neural network specialised for sequences.

Robotics

Machines that sense, decide and act in the physical world, often with AI as the decision layer.

Robustness

Performance maintained under stress: noisy, unexpected, adversarial or shifting inputs.

Robustness (OECD)

AI must function robustly, securely, safely; achieved via traceability and a risk management approach.

Root cause complexity

Difficulty pinning down failure causes when decisions span multiple platforms and data sources.

Ryan Calo taxonomy

Privacy harm split into subjective (internal sense of harm) and objective (external, e.g. refusing a loan).

Safety

System avoids physical, psychological or societal harm under intended and foreseeable use.

Sampling bias

Training data skews toward a subset, favouring specific groups and giving unfair outcomes.

Scraping problem

Training data gathered at petabyte scale from digital content, often personal and without end-user consent.

Section 1557 (HHS OCR rule)

Healthcare nondiscrimination rule requiring covered entities to proactively identify and address discriminatory AI impacts.

Secure multi-party computation

A PET for targeted use; fine for simple arithmetic, compute-intensive for multiplication and division.

Security risk

AI risk including adversarial attacks, hallucinations, deepfakes, data poisoning, overreliance and false sense of security.

Semi-structured data

Partially structured data using tags, elements or markers (e.g. email, XML).

Semi-supervised learning

A small labelled set steers learning over a large unlabelled set; course files LLMs here.

Seven ethical issues

Lawfulness, safety, bias protection, transparency, choice, human intervention, security.

Small language models

LMs with millions to several billion parameters, efficient and often specialised.

Small language models (SLM)

Compact models with far fewer parameters: cheaper, faster, can run on-device.

Snapshots

Saved states of the algorithm and its outputs enabling rollback and change comparison.

Sociotechnical Harms taxonomy

Five themes: representational, allocative, quality-of-service, interpersonal, social system/societal.

Special categories of data

Sensitive personal data needing extra protection under the GDPR and Brazil's LGPD - eight types.

SR 11-7

Federal Reserve standard on model risk management, awaiting AI interpretation.

Stakeholder mapping

Risk strategy ensuring the correct parties are in the decision-making process and spotting risks early.

Static data

Data that does not change, e.g. records of past sales.

Streaming data

Data that changes frequently, e.g. customer visits to a website updating with each visit.

Stress test

Simulating extreme scenarios to evaluate performance and stability under unexpected loads or inputs.

Strict liability

No-fault regime: prove only that the product was defective and the defect caused the harm.

Structured data

Data in fixed fields (rows and columns); easier to analyse, used for business intelligence.

Supervised learning

Training on labelled input-output pairs (classification and regression).

SVM (Support Vector Machine)

Algorithm for classification and regression, used mostly for classification.

SVR (Support Vector Regression)

Algorithm that most commonly produces continuous values.

Synthetic data

Artificially generated data mimicking real data; useful for privacy/scarcity but can inherit bias.

SyntraHome case

AI thermostat overheating traced to a third-party model - stresses vendor screening and pilot deployments.

System

The full operational environment: data, algorithms, models, interfaces and infrastructure.

System card

Documentation of a whole AI system: multiple models/components and their end-to-end behaviour.

Systemic risk

EU category for very large GPAI models above computing thresholds with wide impact, carrying extra duties.

TAKE IT DOWN Act

US law targeting AI-generated nonconsensual and deepfake images; 48-hour removal-request response required.

Tasks and output

OECD dimension covering tasks performed, outputs, action autonomy and evaluation methods.

Temporal bias

Models trained on current data fail to adapt to future changes, so accuracy decays over time.

TEVV

Test, Evaluation, Verification, Validation - cycles built into the deployment timeline.

Texas TRAIGA (HB 149)

Texas Responsible AI Governance Act (1 Jan 2026).

Thaler v Vidal

Fed. Cir. 2022 (cert denied 2023): only humans can be named inventors on a US patent.

Third-party AI

AI acquired from a vendor; deployers retain responsibility to assess risk in their own use context.

Threat modelling

Identifying, understanding, addressing and communicating security risks via structured methods and diagrams.

Three foundational controls

Ethical principles, an oversight body, and policies & procedures (with effectiveness metrics).

Three lines of defence (3LOD)

Management implements (Line 1), risk teams spot (Line 2), internal audit checks (Line 3) - Do: Watch: Check.

Tier 1 foundational guardrails

Guardrails all AI systems need (privacy, transparency, explainability, security, safety) per ISO/IEC 42001 and NIST AI RMF.

Tier 2 risk-based guardrails

Guardrails sized to the use-case risk, from light (retail bot) to rigorous (banking disputes bot).

Tier 3 societal guardrails

Guardrails for impacts on communities, industries and the environment, including emergency shutdown and public policy engagement.

Tiered penalties

Enforcement logic where the highest fines target prohibited/systemic risks, with proportionate caps for SMEs and startups.

Training data

The dataset the model learns from; must be representative, fair and compliant.

Transfer learning model

Reuses knowledge from one task as a starting point for a related task.

Transformer model

Architecture using self-attention to weigh a whole sequence at once; backbone of modern LLMs.

Transformer models

DL that learns context by tracking relationships in sequential data and processing inputs in parallel.

Transparency

Making information available to stakeholders: a system's existence, capabilities, data, logic and limits.

Transparency (privacy principle)

Processing must be clear and accessible to individuals in plain language; duties pervade the EU AI Act and data-protection law.

Transparency threshold rule

Near-universal requirement to disclose that AI is in place, treated by the FTC as a threshold requirement.

Trustworthy AI

AI that is lawful, ethical and robust: valid, safe, fair, accountable, transparent, privacy-protective.

Trustworthy AI characteristics

Seven NIST qualities, valid & reliable, safe, secure & resilient, explainable & interpretable, privacy-enhanced, fair, accountable & transparent.

Turing test

Benchmark: if a human cannot reliably tell a machine from a person in conversation, it shows intelligence.

Unacceptable risk

Banned AI uses such as social credit scoring and real-time remote facial recognition in public spaces.

Underfitting

Model too simple to capture the pattern; poor on training AND new data.

UNESCO Recommendation on the Ethics of AI

A global ethics-of-AI principles document.

Unstructured data

Data with no specific structure (social posts, video, audio, images) that fuels GenAI.

Unsupervised learning

Training on unlabelled data to surface hidden structure (clustering, association).

Use case assessment

Structured process evaluating the viability, risks and ethical implications of applying AI to a specific problem; follows map/measure/manage.

Use case evaluation

Risk strategy determining whether the need warrants AI at all and informing the model type.

User

Role that interacts with AI, recognises when doing so, gives feedback and exercises rights such as notice and human review.

Utah SB 149 / SB 226

Utah AI Policy Act (disclosure in regulated professions) and AI consumer-protection amendments.

Validation data

Held-out data to tune the model and check generalisation before final testing.

Variables

Measurable attributes in the data used as model inputs or predicted outputs.

Variance

Sensitivity of outputs to fluctuations in the training data; high variance is overfitting.

Vendor agreement checklist

Eight areas to evaluate before signing: data, security/safety, bias, product type, technical specs, performance, monitoring, terms of use.

Watermarking

Embedding detectable markers in AI-generated content to signal synthetic origin and provenance.

Weights

Learned connection strengths deciding each input's influence; the core kind of parameter.

Sources

• Current AIGP certification page
• IAPP certification FAQs