AIGP Glossary
Over 260 terms from the BoK, the lecture notes, the EU AI Act and the exam trap notes, in plain English, grouped by topic and searchable.
Start typing and pick a suggestion, or search part of a definition. Try a word such as bias, deployer or drift.
Showing all 269 terms.
No terms match that search. Try a shorter word, or check the spelling.
How to read this. Definitions are simplified for beginners. Where a term is legal or technical, check the exact wording in the source before you rely on it. This is not legal advice.
Source tags. BoK means the AIGP Body of Knowledge v2.1, effective 2 February 2026. Notes D1 to D4 mean the four lecture-note sets for Domains I to IV. AI Act means Regulation (EU) 2024/1689, with article numbers. Trap notes means the exam trap notes. General means a widely used term with no single source. Practice question means a question from the practice screenshots. Community tip means an unverified tip from online discussion.
AI and machine learning basics
The vocabulary you will meet first. Read this section before the others if the words are new. (17 terms)
- Artificial intelligence (AI)
- An engineered system that uses computational techniques to automate tasks. There is no single agreed definition, so note which one your course uses.
- Source: Notes D1
- Algorithm
- A set of instructions and rules a computer follows to complete a task, such as 'if this, then that'.
- Source: Notes D1
- Machine learning (ML)
- A part of AI in which a system learns patterns from data and applies them to new data, rather than following only fixed hand-written rules.
- Source: Notes D1
- Corpus
- A large collection of text or data that an AI system uses to find patterns and make predictions.
- Source: Notes D1
- Training data
- The data used to teach an ML model. The model looks for patterns and relationships in it.
- Source: AI Act Art. 3(29)
- Input data
- Data given to an AI system, or collected directly by it, that the system uses to produce an output.
- Source: AI Act Art. 3(33)
- Labelled data
- Data tagged with the answer or category the model should learn, such as 'spam' or 'not spam'.
- Source: Notes D1
- Model (ML model)
- The learned representation of patterns and relationships that training produces. It is one part of a wider AI system.
- Source: Notes D1
- Inference
- The output a model produces from new input, such as a prediction, a decision or a piece of content.
- Source: Notes D1
- Parameter
- An internal value a model adjusts during training. The values that define how a model behaves are its parameters.
- Source: Notes D1
- Weights
- A type of parameter that sets how strongly connected nodes influence each other in a neural network.
- Source: Notes D1
- Hyperparameter
- A setting chosen before training that controls how learning happens. It is set by people, not learned from the data.
- Source: Notes D1
- Features
- The input variables or attributes a model uses to make predictions, such as income or age in a credit model.
- Source: Notes D1
- Socio-technical system
- A view that people shape technology and technology shapes people, so governance has to cover the social context as well as the code.
- Source: Notes D1
- Artificial narrow intelligence (ANI)
- AI that performs one function within narrow limits. Most systems in use today are generally described this way.
- Source: Notes D1
- Artificial general intelligence (AGI)
- Human-level, general capability across many tasks. The course notes say it does not currently exist.
- Source: Notes D1
- Artificial super intelligence (ASI)
- Capability far beyond human ability. This is a hypothetical concept, not a current system.
- Source: Notes D1
Types of AI and techniques
How different kinds of AI learn, and what they produce. (32 terms)
- Supervised learning
- Learning from labelled data. The examples are labelled before training, and the model learns to predict the label for new data.
- Source: Notes D1
- Unsupervised learning
- Learning from unlabelled data to find structure, such as groups, outliers or relationships. Useful for exploring data.
- Source: Notes D1
- Semi-supervised learning
- A mix of a small amount of labelled data and a large amount of unlabelled data.
- Source: Notes D1
- Reinforcement learning
- A model learns by acting, receiving rewards or penalties, and adjusting its actions to get more reward over time.
- Source: Notes D1
- Reinforcement learning from human feedback (RLHF)
- A method where people rate or compare model outputs and the model is adjusted toward the outputs people prefer.
- Source: Notes D1
- Classification
- A supervised task that predicts a category, such as spam or not spam, or cat or dog.
- Source: Notes D1
- Regression
- A supervised task that predicts a continuous number, such as a house price or a fuel efficiency figure.
- Source: Notes D1
- Clustering
- An unsupervised method that groups similar data points together without being told the groups in advance.
- Source: Notes D1
- Association rule learning
- A data mining technique that finds items that often appear together, such as products bought on the same shopping trip.
- Source: Notes D1
- Decision tree
- A branching, question-and-answer model. Often easier to explain than other models, but sensitive to changes in the data.
- Source: Notes D1
- Neural network
- A model loosely inspired by the brain, made of connected nodes arranged in layers. It can capture complex, non-linear patterns.
- Source: Notes D1
- Deep learning
- A neural network with more than three layers. Its internal reasoning is hard to trace, which is why it is often called a black box.
- Source: Notes D1
- Discriminative model
- A model that classifies an input by learning a boundary between categories.
- Source: Notes D1
- Generative model
- A model that learns the characteristics of data and produces new content with those characteristics.
- Source: Notes D1
- Generative AI
- AI that creates new content such as text, code, images, audio or video. It is usually built on deep learning and large datasets.
- Source: Notes D1
- Foundation model
- A large model trained on very large amounts of data that can be reused and adapted for many related tasks.
- Source: Notes D1
- Transfer learning
- Applying what a model has learned on one task to a different but related task.
- Source: Notes D1
- Fine-tuning
- Further training a pre-trained model on specific data so it performs better for a particular purpose.
- Source: Notes D1
- Large language model (LLM)
- A foundation model built with deep learning on very large amounts of text, designed to understand and generate language.
- Source: Notes D1
- GPT (Generative Pre-trained Transformer)
- A type of neural network architecture that uses an attention technique to work out how words relate to each other in context.
- Source: Notes D1
- Multimodal model
- A model that can take in or produce more than one type of data, such as text, images, speech and video.
- Source: Notes D1
- Small language model (SLM)
- A language model with far fewer parameters than an LLM. Usually smaller and cheaper to run, and easier to host on-premise.
- Source: Notes D1
- Expert system
- An older rules-based system with a knowledge base, an inference engine and a user interface. It supports people rather than replacing them.
- Source: Notes D1
- Computer vision
- AI that recognises objects, faces or patterns in images and video.
- Source: Notes D1
- Natural language processing (NLP)
- AI that helps computers understand, interpret and generate human language, including speech.
- Source: Notes D1
- Robotic process automation (RPA)
- Software 'robots' that automate repetitive digital tasks such as data entry or form processing.
- Source: Notes D1
- Agentic AI
- Systems that take multi-step actions, often by using tools, rather than only producing text in response to a prompt.
- Source: BoK (introduction)
- Prompt
- The instruction or input a user gives a generative AI system to produce an output.
- Source: Notes D4
- Prompt engineering
- Deliberately structuring a prompt, with clear instructions, examples and sequence, to get a specific kind of output.
- Source: Notes D4
- Retrieval-augmented generation (RAG)
- A method that gives a language model relevant reference material from a database when it answers, so the output can draw on data that was not in its training.
- Source: Notes D4
- Synthetic data
- Artificial data that mimics the statistical properties of real data. It is often used to reduce privacy risk.
- Source: Notes D3
- Watermarking
- Embedding a hidden, identifiable signal in AI-generated content so that its artificial origin can be checked.
- Source: Notes D4
Data and privacy
Where data comes from, what counts as personal data, and how it is protected. (38 terms)
- Personal data
- Information relating to an identified or identifiable person. The GDPR definition in Article 4(1) applies, and the AI Act refers to it.
- Source: AI Act Art. 3(50)
- Special category data
- Sensitive personal data, such as health, biometric or racial or ethnic origin data. GDPR Article 9 sets stricter rules and lists its exceptions.
- Source: Notes D2
- Biometric data
- Personal data from technical processing of physical, physiological or behavioural features, such as facial images or fingerprint data.
- Source: AI Act Art. 3(34)
- Pseudonymised data
- Data where direct identifiers are replaced, but the person can still be linked back using extra information. Under the GDPR it is still personal data.
- Source: Notes D2
- Anonymised data
- Data that cannot reasonably be linked back to a person. The GDPR threshold for this is very high, so check it carefully.
- Source: Notes D2
- De-identification
- Removing or masking direct identifiers. Check whether the result still counts as personal data.
- Source: Notes D3
- Data minimisation
- Collecting only what is needed for a stated purpose, and keeping it only as long as it is needed.
- Source: Notes D2
- Purpose limitation
- Using data only for the purpose it was collected for, not for other uses without a separate legal basis.
- Source: Notes D2
- Secondary use
- Using data collected for one purpose for a different purpose.
- Source: Notes D1
- Data persistence
- Data that outlives the person it describes, so retention and deletion rules are needed.
- Source: Notes D1
- Aggregation
- Combining de-identified data from several sources, which can make it possible to re-identify people.
- Source: Notes D1
- Inference (as a privacy risk)
- Working out new facts about a person from existing data, which can reveal things they did not share.
- Source: Notes D1
- Spillover data
- Data collected incidentally, beyond the intended target of a collection.
- Source: Notes D1
- Fair Information Practices (FIPs)
- A set of principles for handling personal data fairly. Common ones include notice, consent, purpose specification, data minimisation and accountability. Do not confuse FIPs with FIPPs or FIPS, which are different terms.
- Source: Notes D2
- Privacy by design
- Building privacy and data protection into a system from the start and throughout its development, rather than adding it later. Ann Cavoukian developed the idea.
- Source: Notes D2
- Privacy by default
- The strictest privacy settings apply automatically, so the user does not have to switch them on.
- Source: Notes D2
- Data protection impact assessment (DPIA)
- A structured assessment of privacy risks for a system or activity. GDPR Article 35 requires one in some cases, usually before deployment.
- Source: Notes D2
- Privacy impact assessment (PIA)
- A tool for identifying, assessing and reducing privacy risk in a project, system or programme.
- Source: Notes D2
- Consent (under the GDPR)
- Must be explicit, voluntary, informed and unambiguous, and the person should be able to opt out. The request should be in plain language.
- Source: Notes D2
- Legal basis
- A lawful reason for processing personal data under GDPR Article 6, such as consent, contract or legitimate interest.
- Source: Notes D2
- Legitimate interest
- A legal basis that depends on a clear purpose, a necessity test and a balancing test against the individual's rights.
- Source: Notes D2
- Automated decision-making (ADM)
- A decision made in whole or part by technology without human involvement. Not every AI system makes decisions, and not every ADM system uses AI.
- Source: Notes D2
- Right to human intervention
- Under GDPR Article 22, people have a right not to be subject to certain solely automated decisions, subject to exceptions, and can ask for human review.
- Source: Notes D2
- Data governance
- Managing data across its life cycle, from collection to sharing and disposal, through policies, roles and standards.
- Source: Notes D2
- Data disposition
- The policy on how long data is kept and how it is securely deleted or destroyed.
- Source: Notes D2
- Data provenance
- Where data came from, who owns it and under what rights it may be used.
- Source: Notes D3
- Data lineage
- The record of how data was transformed and moved through a pipeline over time.
- Source: Notes D3
- Differential privacy
- Adding statistical noise to data or outputs so individuals are harder to pick out. More noise means more privacy and less accuracy.
- Source: Notes D3
- Federated learning
- Training a shared model across several parties' data without pooling the raw data in one place.
- Source: Notes D3
- Homomorphic encryption
- A method that allows calculations on encrypted data without decrypting it first.
- Source: Notes D3
- Secure multi-party computation
- A method where several parties compute a shared result without revealing their own inputs to each other.
- Source: Notes D3
- Trusted execution environment (TEE)
- A secure area of a processor that protects the confidentiality and integrity of code and data while it runs.
- Source: Notes D1
- Intellectual property (IP)
- Legal rights over creations. The main types are patents, trademarks, copyright and trade secrets.
- Source: Notes D2
- Copyright
- Protects original creative and literary works. Whether AI outputs qualify, and whether copyrighted works may be used to train AI, are open legal questions.
- Source: Notes D2
- Fair use
- A US doctrine that allows limited use of copyrighted work without permission. It is decided case by case, so do not treat it as a fixed rule.
- Source: Notes D2
- Derivative work
- A new work based on an existing one. Creating one without permission can raise copyright questions.
- Source: Notes D2
- Licence (model or data)
- The terms that set what you may do with a model or dataset, including whether it may be used for training.
- Source: Notes D2
- Indemnification
- A contract promise to cover the other party's losses, often in AI licences for third-party IP claims. Read the exact scope.
- Source: Notes D2
Testing, performance and documentation
How a model is measured, tested and described before and after release. (30 terms)
- Interpretability
- How well a person can understand a model's internal workings. It is designed in from the start, so the model can be understood.
- Source: Notes D1
- Explainability
- Giving reasons for a specific output after it has been produced. Sources use the term differently, so use your course's wording.
- Source: Notes D1
- Opacity (black box)
- Knowing a model's inputs and outputs but not the reasoning behind a specific decision.
- Source: Notes D1
- Deterministic system
- Gives the same output every time for the same input, because it follows fixed rules.
- Source: Notes D1
- Probabilistic system
- Can give different outputs for similar inputs, because it involves randomness or uncertainty.
- Source: Notes D1
- Accuracy
- The share of outputs that are correct. It can mislead when one outcome is rare: a fraud model that always says 'not fraud' scores 99% if 1% of cases are fraud, yet catches none.
- Source: Notes D3
- Precision
- Of all the cases the model labels positive, the share that really are positive.
- Source: Notes D3
- Recall
- Of all the real positive cases, the share the model finds.
- Source: Notes D3
- F1 score
- A single figure that combines precision and recall. Useful when you need both to be reasonably good.
- Source: Notes D3
- Training, validation and testing data
- Training data teaches the model. Validation data tunes it during development, like a practice quiz. Testing data gives a final, independent check.
- Source: Notes D3
- Overfitting
- The model learns the training data too closely and fails on new data.
- Source: Notes D3
- Underfitting
- The model is too simple to capture the pattern, so it performs poorly, even on the training data.
- Source: Notes D3
- Ground truth
- Verified, known answers used as the reference for measuring performance.
- Source: Notes D3
- Benchmark
- A standard test used to evaluate and compare model performance. The notes name MMLU and MMMU as examples.
- Source: Notes D3
- Pre-deployment pilot
- A trial in a limited setting before full release, to catch problems that testing did not reveal.
- Source: Notes D3
- TEVV (test, evaluate, verify, validate)
- Checks on whether a system works as intended (test), how well it performs (evaluate), whether it was built correctly (verify), and whether it delivers intended outcomes in practice (validate). Definitions vary, so use your course's wording.
- Source: Notes D3
- Repeatability
- The same team getting the same results when it repeats a test under the same conditions.
- Source: Notes D3
- Adversarial testing
- Testing a model with malicious or deceptive inputs to see whether its output can be corrupted.
- Source: Notes D3
- Red teaming
- Testing a system by deliberately trying to make it fail or misbehave, often by simulating attacks.
- Source: Notes D4
- Stress testing
- Simulating extreme conditions to see whether a system breaks and how robust it is.
- Source: Notes D4
- Challenger model
- A new model tested against the current model to see whether it performs better.
- Source: Notes D4
- Model decay
- A model trained on old, static data becomes less accurate over time unless it is monitored and retrained.
- Source: Notes D4
- Data drift
- A change in the inputs a model receives over time, such as new customer slang or new sensor readings.
- Source: Notes D1
- Concept drift
- A change in the relationship between inputs and the outcome, such as new fraud patterns that change what counts as fraud.
- Source: Trap notes
- Hallucination
- Generative AI output that appears true but is false, contradictory or misleading.
- Source: Notes D1
- Brittleness
- A system that performs well under the conditions it was tested on, but fails when conditions shift slightly.
- Source: BoK III.C
- Model card
- Documentation of a model covering its intended use, inputs and outputs, performance and known limits.
- Source: Notes D3
- System card
- Documentation explaining how a group of models works together as one system.
- Source: Notes D4
- Datasheet for datasets
- Documentation of a dataset's purpose, composition, collection process and recommended uses.
- Source: Notes D3
- Confusion matrix
- A table that counts correct and incorrect predictions for each class. Precision, recall and accuracy can all be calculated from it.
- Source: General
Bias and harms
How AI can treat people unfairly, and the wider damage it can cause. (25 terms)
- Bias
- A preference or inclination that gets in the way of impartial judgement. It can lead to unfair outcomes and risks to people's rights.
- Source: Notes D1
- Algorithmic bias
- Systematic, repeatable errors built into a model's design or logic, often caused by unrepresentative training data or flawed modelling choices.
- Source: Notes D1
- Computational (machine) bias
- Errors that come from data collection, sampling, pre-processing or the limits of the algorithm itself.
- Source: Notes D1
- Cognitive bias
- Human thinking shortcuts that distort judgement, such as confirmation bias (favouring what you already believe) or anchoring (leaning too hard on the first figure you see).
- Source: Notes D1
- Societal bias
- Prejudice or favouritism in society that gets reflected in data, access or treatment.
- Source: Notes D1
- Implicit bias
- Unconscious associations about a social group that can influence behaviour and lead to stereotyping.
- Source: Notes D1
- Sampling bias
- When the data does not represent the population the system will be used on.
- Source: Notes D1
- Temporal bias
- When a model does not keep working consistently over time or across changing conditions.
- Source: Notes D1
- Automation bias
- Over-reliance on an automated output, so the human reviewer stops checking it.
- Source: Notes D1
- AI exceptionalism
- The assumption that computer systems are infallible and better than humans, which makes staff less likely to challenge outputs.
- Source: Notes D1
- Harms to individuals
- Harms to civil rights, economic opportunity and safety, including unfair results in hiring, credit, housing, insurance and education.
- Source: Notes D1
- Harms to groups
- Collective harms, such as facial recognition that performs worse for some groups, or surveillance that affects protected groups more.
- Source: Notes D1
- Harms to society
- Wider effects on democratic processes, trust in institutions, access to public services and employment.
- Source: Notes D1
- Harms to companies
- Reputational, cultural, economic and legal harms to an organisation, plus acceleration risks from moving too fast.
- Source: Notes D1
- Representational harm
- A system reinforces unjust stereotypes or societal biases.
- Source: Notes D1
- Allocative harm
- A system distributes resources or opportunities unfairly, for example through an employment algorithm.
- Source: Notes D1
- Quality-of-service harm
- A system works noticeably worse for some social groups, such as voice assistants that struggle with certain accents.
- Source: Notes D1
- Disinformation
- Deliberately deceptive or misleading information, spread to confuse or manipulate.
- Source: Notes D1
- Misinformation
- Incorrect or misleading information that may be shared without intending harm.
- Source: Notes D1
- Deepfake
- Audio, images or video that have been generated or manipulated with AI, often to deceive.
- Source: Notes D1
- Echo chamber and filter bubble
- A situation where people mostly see content that matches their views and little that challenges them.
- Source: Notes D1
- Profiling
- Building a picture of a person's behaviour, preferences or habits from data, often to predict or target them.
- Source: Notes D1
- Sociotechnical harm
- Harm that comes from how a system interacts with its social setting, not only from the code.
- Source: Notes D1
- Acceleration risk
- The risk that fast deployment at scale outpaces anyone's ability to foresee problems, summed up as 'move fast and break things'.
- Source: Notes D1
- Power concentration
- Market or political power held by a few, which can erode freedoms, for example through monopoly or market manipulation.
- Source: Notes D1
Security and attacks
Ways AI systems and their data can be attacked, leaked or misused. (17 terms)
- Adversarial attack
- A deliberately crafted input that makes a model malfunction or produce a wrong or dangerous output.
- Source: Notes D1
- Data poisoning
- Changing training data on purpose to damage a model's performance or behaviour.
- Source: Notes D1
- Membership inference
- An attack that tries to work out whether a particular person's record was used to train a model.
- Source: Notes D1
- Model extraction (theft)
- Repeatedly querying a model to copy its behaviour or recreate it without permission.
- Source: Notes D1
- Model inversion
- An attack that uses a model's outputs to recover information about the data it was trained on.
- Source: Notes D1
- Model evasion
- Crafting inputs that slip past a model's detection, such as spam or malware filters.
- Source: Notes D1
- Prompt injection
- Hiding instructions inside content, such as a web page or document, so an AI system follows them without the user intending it.
- Source: Notes D1
- Data leakage
- Data exposed by accident, for example through human error or a misconfigured system.
- Source: Notes D1
- Data loss
- Data that can no longer be recovered, for example after a lost laptop, ransomware or deletion.
- Source: Notes D1
- False sense of security
- Believing an AI system is objective, unbiased or always correct, when it may not be.
- Source: Notes D1
- Misuse
- Using a system in a way it was not designed or permitted for, including malicious use.
- Source: Notes D1
- Threat modelling
- A structured way to identify what could go wrong, rank the threats, choose countermeasures and check the result. It asks what we are working on, what can go wrong, what we will do, and whether we did a good job.
- Source: Notes D3
- AI incident
- An event where an AI system causes harm, malfunctions or nearly does. Incidents often arise from brittleness, poor data, weak testing or drift.
- Source: BoK III.C
- Incident response
- A structured process for handling incidents in six stages: preparation, identification, containment, eradication, recovery and lessons learned.
- Source: Notes D4
- AI bill of materials (AI-BOM)
- An inventory of a system's components, such as data, endpoints, APIs, model artefacts and downstream dependencies.
- Source: Notes D4
- Shadow AI
- AI used inside an organisation without approval, tracking or governance.
- Source: General
- Least privilege
- Giving a system or person only the access needed for the task.
- Source: General
Ethics and responsible AI
The values behind AI governance, and the words used to describe them. (18 terms)
- Ethical AI
- The values, or compass: the principles that guide what AI should and should not do.
- Source: Notes D1
- Responsible AI
- The action: how ethical principles are put into practice through processes, roles and controls.
- Source: Notes D1
- Trustworthy AI
- The outcome: AI that people have good reason to trust. The three terms are often used interchangeably, so check which sense is meant.
- Source: Notes D1
- Accountability
- Clear responsibility for a system's behaviour, with its actions, decisions and outcomes traceable.
- Source: Notes D1
- Contestability
- People can question, challenge and appeal an AI output or decision.
- Source: Notes D1
- Fairness
- Standards set for each system, applied consistently so that individuals and groups are treated fairly.
- Source: Notes D1
- Transparency
- Being open about when AI is used, what it is for and how decisions are made, to a degree suited to the audience.
- Source: BoK I.A
- Safety
- Minimising harm from a system. Notes describe AI safety as minimising harm.
- Source: Notes D1
- Robustness
- A system keeps working as intended when inputs are unexpected or conditions are difficult.
- Source: BoK III.C
- Reliability
- A system performs consistently and as intended over time.
- Source: BoK I.A
- Alignment
- An AI system pursues goals that match what its operators intend, their preferences and ethical principles.
- Source: Notes D1
- Misalignment
- A mismatch between what a system is meant to do and what it actually does. Inner misalignment is between programmed goals and behaviour; outer misalignment is between human intentions and coded objectives.
- Source: Notes D1
- Intended, specified and emergent goals
- Intended goals are what the operator wants, specified goals are what is programmed, and emergent goals are what the system actually pursues.
- Source: Notes D1
- Human-in-the-loop
- A person reviews or approves an output or decision before it takes effect.
- Source: Notes D1
- Human-on-the-loop
- A person monitors the system and can step in, but does not approve each decision.
- Source: Notes D1
- Human-out-of-the-loop
- The system acts without a person involved in each decision.
- Source: Notes D1
- Human oversight
- People with the knowledge, time and authority to understand, question and override a system. The AI Act sets specific duties in Article 14.
- Source: AI Act Art. 14
- Ethics by design
- Identifying and resolving ethical issues during the plan and design stage, not after launch. Its principles include respect for human agency, privacy, fairness, well-being, transparency and accountability.
- Source: Notes D1
Governance, risk and operations
Policies, assessments and controls across the AI life cycle. (25 terms)
- AI governance
- The policies, roles, oversight and processes that guide how an organisation builds, buys, uses and retires AI across its life cycle.
- Source: BoK I
- Governance vs compliance
- Governance sets the house rules and how decisions are made. Compliance checks that the rules, and the law, are being followed.
- Source: General
- Framework vs standard vs law
- A framework is structured guidance, often voluntary. A standard is a published specification, and some can be certified against. A law is binding. Check the status of each source you rely on.
- Source: BoK II.D
- Risk management
- Identifying, assessing, treating and monitoring risks throughout the AI life cycle.
- Source: BoK III.A
- Risk tolerance
- How much risk an organisation is willing to accept, which shapes the controls it needs.
- Source: BoK I.B
- Harms matrix
- A grid that scores how likely a harm is against how severe it would be, so risks can be ranked.
- Source: Notes D3
- Risk mitigation hierarchy
- The order for dealing with a risk: avoid or remove it where you can, then reduce it, then manage what remains. Used alongside the harms matrix.
- Source: Notes D3
- Residual risk
- The risk that remains after controls have been applied.
- Source: Notes D4
- Controls
- Measures that reduce risk, such as policies, technical safeguards, human review and training.
- Source: Notes D3
- Impact assessment (AIIA)
- A structured review of what could go wrong for people and the organisation. It is usually done before a system is built or deployed, and the first one often belongs in the plan and design stage.
- Source: Notes D1
- Conformity assessment
- A set of technical and non-technical checks, plus documentation, showing that a high-risk AI system meets the EU AI Act requirements before it is placed on the market.
- Source: AI Act Art. 3(20), 43
- CE marking
- A marking a provider uses to show that its AI system meets the EU requirements that apply to it.
- Source: AI Act Art. 3(24)
- Notified body
- An independent body designated to carry out third-party conformity assessments under the EU AI Act.
- Source: AI Act Art. 3(22)
- Post-market monitoring
- Collecting and reviewing data on a system after release, to spot problems and feed them back into governance.
- Source: AI Act Art. 3(25), 72
- Serious incident
- Under the EU AI Act, an incident or malfunction that leads to specified serious harms. Check the definition in Article 3(49) and the reporting duty in Article 73.
- Source: AI Act Art. 3(49), 73
- Technical documentation
- Records showing how a system was designed, built, tested and monitored. Used to show compliance and to support oversight.
- Source: AI Act Art. 11
- Instructions for use
- Information a provider gives a deployer on how to use a system as intended and safely.
- Source: AI Act Art. 3(15)
- Audit vs assessment
- An assessment looks forward at possible risks. An audit looks back to check whether controls were actually followed.
- Source: Trap notes
- Third-party risk
- Risk from vendors, suppliers, data or models that the organisation did not build, managed through procurement, contracts and monitoring.
- Source: BoK I.C
- Acceptable use policy
- The rules on how staff and users may use an AI tool, including what they must not put into it.
- Source: Notes D3
- Vendor or licence agreement
- The contract that sets what a vendor provides and what you may do with a model or data. Key terms and risks should be identified before signing.
- Source: BoK IV.B
- Change management
- A controlled process for approving, testing and recording changes to a system.
- Source: Notes D4
- Decommissioning
- Retiring an AI system safely: deciding what to archive, delete or keep, and communicating with users and third parties.
- Source: Notes D4
- Deactivate or localise
- A control that lets you switch off or restrict an AI system, for example for regulatory reasons or because it is performing badly.
- Source: BoK IV.C
- AI literacy
- The skills and understanding that staff need to use and oversee AI appropriately. Article 4 of the EU AI Act requires providers and deployers to take measures for it.
- Source: AI Act Art. 4
EU AI Act
The EU's AI law: its key terms, roles and risk levels. Check the current consolidated text before relying on any detail. (26 terms)
- EU AI Act
- Regulation (EU) 2024/1689, the EU's risk-based law on artificial intelligence. It has been amended since publication, so check the current consolidated text.
- Source: AI Act
- AI system (EU AI Act definition)
- A machine-based system designed to run with varying levels of autonomy, which may infer outputs such as predictions, content or decisions from its input. Check the full definition in Article 3(1).
- Source: AI Act Art. 3(1)
- General-purpose AI (GPAI) model
- A model trained on large amounts of data, showing significant generality and able to perform a range of tasks. The course notes also call it a foundation model.
- Source: AI Act Art. 3; Notes D2
- GPAI model with systemic risk
- A GPAI model presumed to have high-impact capabilities when its training compute is greater than 10²⁵ floating point operations. The Commission can change the threshold. Check Article 51.
- Source: AI Act Art. 51(2)
- Prohibited practice
- A use the Act bans, listed in Article 5, subject to limited exceptions.
- Source: AI Act Art. 5
- High-risk AI system
- A system that carries the Act's heaviest obligations. Article 6 and Annex III set out which systems count.
- Source: AI Act Art. 6
- Annex III
- The list of use areas that can make a system high-risk. The areas include biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration and democratic processes.
- Source: AI Act Annex III
- Article 6(3) exception
- Some Annex III systems are not high-risk if they only perform a narrow procedural task, improve a completed human activity, detect deviations without replacing human assessment, or prepare an assessment. Profiling always stays high-risk. Check the full text.
- Source: AI Act Art. 6(3)
- Limited-risk (transparency) systems
- Systems with transparency duties, such as telling people they are interacting with an AI system.
- Source: AI Act Art. 50
- Minimal-risk systems
- Low-risk systems that face few or no specific obligations under the Act. Voluntary codes may apply.
- Source: Notes D2
- Provider
- Develops an AI system or GPAI model, or has one developed for them, and places it on the market or puts it into service under its name. The most heavily regulated role.
- Source: AI Act Art. 3(3)
- Deployer
- Uses an AI system under its own authority. Using a tool is usually a deployer role, while building it is a provider role.
- Source: AI Act Art. 3(4)
- Importer
- An EU-based operator that places a system on the EU market when it carries the name or trademark of a provider based outside the EU.
- Source: AI Act Art. 3(6)
- Distributor
- A supply-chain party, other than the provider or importer, that makes a system available on the EU market.
- Source: AI Act Art. 3(7)
- Product manufacturer
- A party that puts an AI system on the market, or uses it with its own product, under its own name or trademark.
- Source: Notes D2
- Authorised representative
- An EU-based person that a non-EU provider has mandated in writing to carry out certain provider obligations.
- Source: AI Act Art. 3(5)
- Operator
- The umbrella term for providers, product manufacturers, deployers, authorised representatives, importers and distributors.
- Source: AI Act Art. 3(8)
- Substantial modification
- A significant change made after a system is placed on the market. A party that makes one may take on provider duties, so check Article 25.
- Source: AI Act Art. 3(23), 25
- Intended purpose
- The use the provider intends for the system, including the context and conditions of use.
- Source: AI Act Art. 3(12)
- Reasonably foreseeable misuse
- Use of a system in a way that does not match its intended purpose but can reasonably be expected to happen.
- Source: AI Act Art. 3(13)
- Fundamental rights impact assessment (FRIA)
- An assessment of how certain high-risk systems affect fundamental rights, which some deployers must complete before use.
- Source: AI Act Art. 27
- EU database registration
- Some high-risk systems must be registered in an EU database. Check Article 49 for who must do it.
- Source: AI Act Art. 49
- Market surveillance authority
- The national authority that checks AI systems on the market for compliance with the Act.
- Source: AI Act Art. 3(26)
- Extraterritorial scope
- The Act can reach beyond EU borders, for example providers placing systems on the EU market from outside the EU, and non-EU operators whose output is used in the EU. Check Article 2.
- Source: Notes D2; AI Act Art. 2
- Scope exceptions
- Some uses fall outside the Act, including those used exclusively for military, defence or national security purposes, purely personal non-professional use, and certain research. Check Article 2 for the full list.
- Source: Notes D2; AI Act Art. 2
- AI Office
- The European Commission's function that supports implementation of the Act and supervises general-purpose AI models.
- Source: AI Act Art. 3(47)
Other laws, standards and frameworks
Privacy, consumer, IP and standards terms that the BoK asks you to know. (23 terms)
- GDPR
- The General Data Protection Regulation (2018), the EU's privacy law. It applies when AI processes personal data.
- Source: Notes D2
- Data protection authority (DPA)
- The national regulator that enforces privacy law such as the GDPR and responds to requests from organisations.
- Source: Notes D2
- CCPA and CPRA (California)
- California privacy laws from 2018 and 2020, which give California residents rights over their personal information.
- Source: Notes D2
- BIPA (Illinois)
- The Biometric Information Privacy Act (2008), which regulates how biometric identifiers are collected and used.
- Source: Notes D2
- Non-discrimination laws
- Laws that prevent unequal treatment, for example in employment, credit, lending, housing and insurance.
- Source: BoK II.B
- Consumer protection law
- Laws against unfair or deceptive acts or practices towards consumers.
- Source: BoK II.B
- FTC (Federal Trade Commission)
- The US consumer protection regulator, which can act against unfair or deceptive practices, including misleading AI claims.
- Source: Notes D2
- ECOA
- The Equal Credit Opportunity Act (US). It prohibits discrimination against credit applicants on protected grounds.
- Source: Notes D2
- FCRA
- The Fair Credit Reporting Act (US). It regulates consumer reporting and requires specific reasons for certain adverse actions.
- Source: Notes D2
- Adverse impact
- A selection procedure that disadvantages a protected group, even if no one intended it. Check hiring tools for it.
- Source: Notes D2
- Product liability
- Rules on who is responsible when a product causes harm, including design and manufacturing defects.
- Source: BoK II.B
- Software as a Medical Device (SaMD)
- Software that performs medical functions without being part of a hardware medical device. It is regulated under healthcare rules.
- Source: Notes D2
- NIST AI RMF
- The US voluntary AI Risk Management Framework, organised around Govern, Map, Measure and Manage. It is not specific to one sector.
- Source: Notes D2
- ISO/IEC 42001
- An international standard for an AI management system, against which an organisation can be certified.
- Source: Notes D2
- ISO/IEC 22989
- A standard that sets out AI concepts and terminology, giving a shared vocabulary.
- Source: Notes D2
- ISO/IEC 42005
- A standard on AI system impact assessment. Check its scope before relying on it.
- Source: BoK II.D
- ISO 31000
- Guidelines for risk management, with principles and a process: identify risks, evaluate probability and severity, then mitigate.
- Source: Notes D2
- IEEE 7000-21
- A standard that addresses ethical concerns in system design.
- Source: Notes D2
- OECD AI Principles
- Values and recommended practices for trustworthy AI, set out by the OECD, an intergovernmental organisation founded in 1961.
- Source: Notes D2
- HUDERAF
- The Council of Europe's Human Rights, Democracy and the Rule of Law Assurance Framework for AI. Check the official text before citing it.
- Source: Notes D2
- Americans with Disabilities Act (ADA)
- US law that prohibits disability discrimination, including in employment, and in many cases requires reasonable accommodation.
- Source: Practice question
- Title VII of the Civil Rights Act of 1964
- US law that prohibits employment discrimination based on race, colour, religion, sex or national origin.
- Source: Practice question
- ADEA (Age Discrimination in Employment Act)
- US law that protects workers aged 40 and over from age discrimination in employment.
- Source: Practice question
Exam and certification terms
How the exam is organised and the words it uses. (18 terms)
- AIGP
- AI Governance Professional, the IAPP certification this site helps you prepare for.
- Source: General
- IAPP
- The International Association of Privacy Professionals, which runs the AIGP exam.
- Source: General
- Body of Knowledge (BoK)
- The IAPP's published outline of what the exam covers. It sets out four domains, their competencies and performance indicators, and the blueprint question ranges. Check the version and effective date.
- Source: BoK
- Domain
- One of the four main areas of the BoK: I, foundations of AI governance; II, laws, standards and frameworks; III, governing AI development; IV, governing AI deployment and use.
- Source: BoK
- Competency
- A broad area of knowledge within a domain, labelled with a code such as I.A or III.B.
- Source: BoK
- Performance indicator
- A specific task or ability that shows you have a competency. Exam questions assess these.
- Source: BoK
- Blueprint
- The table of minimum and maximum questions for each domain in the BoK.
- Source: BoK
- Scored and unscored questions
- The exam includes scored questions that count toward your result and unscored questions that do not. Check the IAPP exam page for the current split.
- Source: General
- Bloom's taxonomy
- A ladder of thinking levels: remember, understand, apply, analyse, evaluate and create. The BoK says most exam questions sit at the remember/understand and apply/analyse levels.
- Source: BoK
- Scenario-based question
- A question that describes a workplace situation and asks which response fits best. The trap notes describe the exam as relying heavily on these.
- Source: Trap notes
- Distractor
- A wrong answer option written to look plausible. Good distractors often use correct terminology in the wrong context.
- Source: Trap notes
- Role misallocation trap
- A question that tempts you to assign a duty to the party with the familiar business name, rather than the party with the legal role.
- Source: Trap notes
- Correct fact, wrong answer
- An option that is accurate as a statement but does not answer the question asked. Re-read the final sentence of the scenario to check the role, jurisdiction and life-cycle stage.
- Source: Trap notes
- Sequencing question
- Asks you to put steps in the right order, such as stages of the AI life cycle or the stages of an incident response.
- Source: Practice question
- Role distinction question
- Asks you to match a duty to the right party, such as a provider, a deployer or an importer.
- Source: Practice question
- Multiple-select question
- Asks you to choose several valid options from a list, such as the required elements of technical documentation.
- Source: Practice question
- Definition or recall question
- Asks you to recall a term, a standard or a legal threshold.
- Source: Practice question
- Best-answer heuristic (community tip)
- A tip often shared online is to choose the option that reduces risk and strengthens accountability when two options look right. Treat it as a starting point and check it against the scenario. It is not an official rule.
- Source: Community tip (unverified)