ISO 22989

0.0(0)

Studied by 0 people

Call Kai

Learn

Practice Test

Spaced Repetition

Match

Flashcards

Knowt Play

Card Sorting

1/131

There's no tags or description

Looks like no tags are added yet.

132 Terms

New cards

★ AI agent [3.1.1]

OFFICIAL: automated entity that senses and responds to its environment and takes actions to achieve its goals HARD WORDS: 'automated' = works by itself without a person doing each step. 'entity' = a thing that exists, here a piece of software. 'senses' = notices what is going on. 'environment' = everything around it. IN SIMPLE WORDS: An AI agent is a program that notices what is happening and then acts by itself to reach a goal. Like a robot vacuum that sees dirt and goes to clean it. REAL EXAMPLE: An AI assistant that reads a new support email, decides it is a password reset, and sends the reset link without a person stepping in. EXAM TIP: Agent = senses, decides and ACTS towards goals. Not every AI system is an agent; a model that only predicts is not.

New cards

AI component [3.1.2]

OFFICIAL: functional element that constructs an AI system HARD WORDS: 'functional element' = a working part. 'constructs' = helps build up. IN SIMPLE WORDS: An AI component is one working part of a bigger AI system. Like one Lego brick in a Lego castle. REAL EXAMPLE: In a fraud-detection system, the model that scores each payment is one AI component; the screen that shows alerts is another part of the same system. EXAM TIP: Component = PART. AI system (3.1.4) = the WHOLE.

New cards

★ artificial intelligence (AI) [3.1.3]

OFFICIAL: (discipline) research and development of mechanisms and applications of AI systems HARD WORDS: 'discipline' = a field of study or work. 'mechanisms' = ways of making things work. 'applications' = real uses. IN SIMPLE WORDS: AI is the whole field of studying and building AI systems. It is a subject, like maths or science, not one product. REAL EXAMPLE: A university AI research group works in the field of AI. The chatbot they build is an AI system. EXAM TIP: In this standard 'AI' means the DISCIPLINE. The thing you build and use is an 'AI system' (3.1.4).

New cards

★ artificial intelligence system (AI system) [3.1.4]

OFFICIAL: engineered system that generates outputs such as content, forecasts, recommendations or decisions for a given set of human-defined objectives HARD WORDS: 'engineered' = designed and built by people. 'outputs' = what it produces. 'forecasts' = guesses about what will happen. 'objectives' = goals. 'human-defined' = set by people. IN SIMPLE WORDS: An AI system is a machine people build to make things like answers, guesses, suggestions or choices. People decide what it is for. Like a sat-nav that suggests the best route. REAL EXAMPLE: A CV-screening tool that ranks job applicants for a recruiter is an AI system: people set the goal (find good candidates) and it outputs a ranked list. EXAM TIP: THE key term: 42001 governs AI systems. Note the goals are set by HUMANS, and systems run at different levels of automation.

New cards

★ autonomy (autonomous) [3.1.5]

OFFICIAL: characteristic of a system that is capable of modifying its intended domain of use or goal without external intervention, control or oversight HARD WORDS: 'characteristic' = a feature. 'modifying' = changing. 'intended domain of use' = the job and place it was meant for. 'external intervention, control or oversight' = someone outside stepping in, steering or watching. IN SIMPLE WORDS: An autonomous system can change its own goal or job without anyone stepping in. That goes much further than just running by itself. Few real systems do this. REAL EXAMPLE: Automated: a bot sends invoices every Friday because it was told to. Autonomous: a bot decides by itself to start chasing a new kind of debt nobody asked it to. Few real systems are truly autonomous. EXAM TIP: Autonomy = can change its OWN GOAL. Automation (3.1.7) = runs set steps without a person. Heteronomy (3.1.16) = under outside control.

New cards

application specific integrated circuit (ASIC) [3.1.6]

OFFICIAL: integrated circuit customized for a particular use HARD WORDS: 'integrated circuit' = a computer chip. 'customized' = made for one purpose. IN SIMPLE WORDS: An ASIC is a computer chip made for one special job. It is not a chip for everything. Like a key cut for just one lock. REAL EXAMPLE: A chip built only to run AI image recognition inside a security camera. EXAM TIP: Shows up in the AI ecosystem (clause 8) as specialised hardware for running AI.

New cards

★ automatic (automation; automated) [3.1.7]

OFFICIAL: pertaining to a process or system that, under specified conditions, functions without human intervention HARD WORDS: 'pertaining to' = about. 'process' = a set of steps. 'under specified conditions' = when things are as expected. 'human intervention' = a person stepping in. IN SIMPLE WORDS: Automated means a system does its steps by itself, as long as things are normal. It follows set rules and does not pick its own goals. Like a dishwasher running its cycle. REAL EXAMPLE: A system that automatically blocks logins from overseas IP addresses at night. Automated means it follows set rules without a person, not that it chooses its own goals. EXAM TIP: Automation ≠ autonomy. Automation follows set rules; autonomy can change its own goals.

New cards

cognitive computing [3.1.8]

OFFICIAL: category of AI systems that enables people and machines to interact more naturally HARD WORDS: 'category' = a type. 'interact' = work and talk with each other. 'naturally' = in a normal, human way. IN SIMPLE WORDS: This kind of AI lets people and machines talk and work together in a normal way. Like chatting to a voice helper instead of clicking menus. REAL EXAMPLE: A voice assistant you can talk to normally to book a meeting, instead of clicking through menus. EXAM TIP: Linked to machine learning, speech, NLP, computer vision and human-machine interfaces.

New cards

★ continuous learning (continual learning; lifelong learning) [3.1.9]

OFFICIAL: incremental training of an AI system that takes place on an ongoing basis during the operation phase of the AI system life cycle HARD WORDS: 'incremental training' = learning a little at a time. 'ongoing basis' = all the time. 'operation phase' = the time when it is being used for real. 'life cycle' = the whole life of the system. IN SIMPLE WORDS: The AI keeps learning from new data while people are using it. It does not stop learning after launch. Like a student who keeps learning on the job. REAL EXAMPLE: A spam filter that updates itself every day from emails users mark as spam. Risk: it can drift, so it needs continuous validation. EXAM TIP: Systems that keep learning need the extra life cycle stage of CONTINUOUS VALIDATION, because they can drift and get worse.

New cards

connectionism (connectionist paradigm; connectionist model; connectionist approach) [3.1.10]

OFFICIAL: form of cognitive modelling that uses a network of interconnected units that generally are simple computational units HARD WORDS: 'cognitive modelling' = copying how thinking works. 'interconnected units' = many small parts linked together. 'computational units' = tiny parts that do sums. IN SIMPLE WORDS: Connectionism builds AI from lots of small linked parts that each do a tiny job. Together they learn, a bit like brain cells. Neural networks work this way. REAL EXAMPLE: A neural network that learns to recognise handwriting through millions of tiny connected calculations, with no written rules. EXAM TIP: Connectionist = subsymbolic AI (3.1.34), the opposite of rule-based symbolic AI (3.1.33).

New cards

data mining [3.1.11]

OFFICIAL: computational process that extracts patterns by analysing quantitative data from different perspectives and dimensions, categorizing them, and summarizing potential relationships and impacts HARD WORDS: 'computational process' = steps a computer follows. 'extracts' = pulls out. 'patterns' = things that repeat. 'quantitative data' = numbers. 'categorizing' = sorting into groups. IN SIMPLE WORDS: Data mining means digging through lots of data to find hidden patterns. Like sorting a huge pile of shop receipts to see what people buy together. REAL EXAMPLE: A retailer mining two years of sales data and finding that sunscreen and hats sell together in October. EXAM TIP: Data mining finds patterns. It is listed as one of the fields of AI in clause 9.3.

New cards

declarative knowledge [3.1.12]

OFFICIAL: knowledge represented by facts, rules and theorems HARD WORDS: 'knowledge' = useful information. 'represented' = written down or stored. 'theorems' = rules that have been proven true. IN SIMPLE WORDS: Declarative knowledge is knowing WHAT is true: facts and rules. It is not about how to do things. Like knowing that water boils at 100 degrees. REAL EXAMPLE: "Essential Eight Maturity Level 2 requires patching critical vulnerabilities within two weeks" is declarative knowledge. EXAM TIP: Declarative = knowing WHAT. Procedural (3.1.28) = knowing HOW. Declarative often has to be turned into procedural before a system can use it.

New cards

★ expert system [3.1.13]

OFFICIAL: AI system that accumulates, combines and encapsulates knowledge provided by a human expert or experts in a specific domain to infer solutions to problems HARD WORDS: 'accumulates' = collects. 'combines' = joins together. 'encapsulates' = packs up. 'knowledge' = useful information. 'domain' = a particular area or subject. 'infer' = work out. IN SIMPLE WORDS: An expert system holds a human expert's know-how as rules. It uses those rules to solve problems in one area. Like a computer helper that asks you questions to find a fault in your car. REAL EXAMPLE: A rule-based tool that asks about a client's controls and suggests their Essential Eight maturity level, built from an auditor's rules. EXAM TIP: Expert systems are SYMBOLIC AI: they use written rules, not learning from data.

New cards

★ general AI (AGI; artificial general intelligence) [3.1.14]

OFFICIAL: type of AI system that addresses a broad range of tasks with a satisfactory level of performance HARD WORDS: 'type' = kind. 'broad range of tasks' = lots of different jobs. 'satisfactory level of performance' = doing them well enough. IN SIMPLE WORDS: General AI could do many different kinds of jobs well, not just one. Some people use it to mean AI that can do anything a person can. It does not exist yet. REAL EXAMPLE: A system that could write contracts, diagnose illness and drive a car equally well. Today's systems are narrow AI, not AGI. EXAM TIP: General AI (AGI) vs narrow AI (3.1.24). Today's real systems are narrow AI.

New cards

genetic algorithm (GA) [3.1.15]

OFFICIAL: algorithm which simulates natural selection by creating and evolving a population of individuals (solutions) for optimization problems HARD WORDS: 'algorithm' = a set of steps for solving a problem. 'simulates natural selection' = copies how nature keeps the strongest. 'population' = a big group. 'optimization problems' = finding the best answer. IN SIMPLE WORDS: A genetic algorithm solves problems by copying evolution. It makes lots of possible answers, keeps the best ones, mixes them, and repeats. Like breeding the fastest racing pigeons each year. REAL EXAMPLE: Finding the best delivery-truck routes by evolving thousands of route plans and keeping the shortest each round. EXAM TIP: A soft computing method (5.8). Good for finding the best answer among very many options.

New cards

★ heteronomy (heteronomous) [3.1.16]

OFFICIAL: characteristic of a system operating under the constraint of external intervention, control or oversight HARD WORDS: 'characteristic' = a feature. 'under the constraint of' = held back by. 'external intervention, control or oversight' = someone outside stepping in, steering or watching. IN SIMPLE WORDS: A heteronomous system works under outside control. People or other systems can watch it, steer it or stop it. Like a learner driver with an instructor who has their own brake. REAL EXAMPLE: A loan-approval AI that suggests decisions, but a credit officer must approve or reject each one. Human oversight makes it heteronomous. EXAM TIP: Heteronomy = OPPOSITE of autonomy. Human oversight makes a system heteronomous, which 42001 generally wants.

New cards

★ inference [3.1.17]

OFFICIAL: reasoning by which conclusions are derived from known premises HARD WORDS: 'reasoning' = thinking something through. 'conclusions' = answers you reach. 'derived' = worked out. 'premises' = things you already know or assume. IN SIMPLE WORDS: Inference means working out an answer from what you already know. In AI it also means the system giving an answer from its model and the new data. Like guessing it rained because the grass is wet. REAL EXAMPLE: The model sees a login from a new country at 3am and infers the account may be compromised. EXAM TIP: 'Inference' means both the thinking and its result. In AI a premise can be a fact, rule, model, feature or raw data.

New cards

internet of things (IoT) [3.1.18]

OFFICIAL: infrastructure of interconnected entities, people, systems and information resources together with services that process and react to information from the physical world and virtual world HARD WORDS: 'infrastructure' = the basic set-up. 'interconnected entities' = linked things. 'information resources' = sources of information. 'process' = handle and use. 'physical world' = real things. 'virtual world' = the online world. IN SIMPLE WORDS: The internet of things is everyday objects linked to the internet. They sense the real world and send information so services can react. Like a smart fridge that tells your phone you are out of milk. REAL EXAMPLE: Smart building sensors that report temperature and turn on air-conditioning when a room gets hot. EXAM TIP: IoT system (3.1.20) = the whole set-up. IoT device (3.1.19) = one sensor or actuator in it.

New cards

IoT device [3.1.19]

OFFICIAL: entity of an IoT system that interacts and communicates with the physical world through sensing or actuating HARD WORDS: 'entity' = a thing. 'IoT system' = the whole linked set-up of devices and services. 'sensing' = noticing or measuring. 'actuating' = making something move or change. IN SIMPLE WORDS: An IoT device is one gadget in an IoT set-up. It either senses something or does something in the real world. Like a smoke alarm (senses) or a smart lock (does). REAL EXAMPLE: A door sensor (senses) or a smart lock (acts) on an office network. EXAM TIP: Device = a sensor OR an actuator.

New cards

IoT system [3.1.20]

OFFICIAL: system providing functionalities of IoT HARD WORDS: 'functionalities' = the jobs it can do. 'IoT' = everyday objects linked to the internet. IN SIMPLE WORDS: An IoT system is the whole set-up: devices, the gadgets that link them, and the services behind them. Like a home with smart lights, a hub and an app working together. REAL EXAMPLE: A warehouse's full setup of shelf sensors, gateways and the cloud app that tracks stock. EXAM TIP: Can include IoT devices, gateways, sensors and actuators (Note 1).

New cards

★ knowledge [3.1.21]

OFFICIAL: (artificial intelligence) abstracted information about objects, events, concepts or rules, their relationships and properties, organized for goal-oriented systematic use HARD WORDS: 'abstracted information' = information boiled down to what matters. 'objects, events, concepts or rules' = things, happenings, ideas or rules. 'organized for goal-oriented systematic use' = arranged so it can be used, step by step, to reach a goal. IN SIMPLE WORDS: In AI, knowledge is information arranged so a system can use it to reach a goal. It does not mean the system understands anything. Like a well-sorted recipe book. REAL EXAMPLE: A knowledge base of security controls and which threats each one reduces, used by an assistant to recommend controls. EXAM TIP: Knowledge in AI does NOT mean understanding (Note 1). Information = data given context.

New cards

★ life cycle [3.1.22]

OFFICIAL: evolution of a system, product, service, project or other human-made entity, from conception through retirement HARD WORDS: 'evolution' = how it changes over time. 'conception' = the first idea. 'retirement' = being switched off for good. IN SIMPLE WORDS: A life cycle is the whole life of something, from the first idea to the end. Like a butterfly going from egg to caterpillar to butterfly. REAL EXAMPLE: A chatbot's life cycle: the idea (2024), building it, testing, going live, monitoring, re-evaluating, and finally retiring it in 2027. EXAM TIP: Clause 6 gives the AI system life cycle stages: inception, design and development, verification and validation, deployment, operation and monitoring, continuous validation, re-evaluation, retirement.

New cards

★ model [3.1.23]

OFFICIAL: physical, mathematical or otherwise logical representation of a system, entity, phenomenon, process or data HARD WORDS: 'physical, mathematical or otherwise logical representation' = a copy made of real things, maths or logic. 'entity' = a thing. 'phenomenon' = something that happens. 'process' = a set of steps. IN SIMPLE WORDS: A model is a simple copy of something real that helps you understand or predict it. Like a globe is a model of the Earth. REAL EXAMPLE: A machine learning model that represents 'what a fraudulent payment looks like', learned from past payments. EXAM TIP: In AI, 'model' usually means a machine learning model (3.3.7) that turns input into a prediction.

New cards

★ narrow AI [3.1.24]

OFFICIAL: type of AI system that is focused on defined tasks to address a specific problem HARD WORDS: 'type' = kind. 'defined tasks' = set jobs. 'specific problem' = one particular problem. IN SIMPLE WORDS: Narrow AI is built for one job or a small set of jobs. It can be very good at that job but useless at others. Like a calculator that only does sums. REAL EXAMPLE: A model that only detects phishing emails. It cannot also write your board report. EXAM TIP: Almost all AI in use today is narrow AI. Opposite: general AI (3.1.14).

New cards

★ performance [3.1.25]

OFFICIAL: measurable result HARD WORDS: 'measurable' = something you can count or check. 'result' = what happens in the end. IN SIMPLE WORDS: Performance is a result you can measure. It can be a number or a judgement. Like a race time or a teacher's grade. REAL EXAMPLE: A fraud model catches 92% of fraud with 1% false alarms. Those figures describe its performance. EXAM TIP: Can be about numbers or judgements. 42001 uses the same definition (3.11).

New cards

planning [3.1.26]

OFFICIAL: (artificial intelligence) computational processes that compose a workflow out of a set of actions, aiming at reaching a specified goal HARD WORDS: 'computational processes' = steps a computer follows. 'compose a workflow' = put steps in order. 'specified goal' = a set target. IN SIMPLE WORDS: In AI, planning means the computer working out a list of steps to reach a goal. Like a sat-nav working out each turn to get you home. REAL EXAMPLE: An AI agent planning the steps to onboard a new staff member: create account, assign laptop, book training. EXAM TIP: In management standards 'planning' can also mean what PEOPLE do (Note 1). Context tells you which.

New cards

★ prediction [3.1.27]

OFFICIAL: primary output of an AI system when provided with input data or information HARD WORDS: 'primary output' = the main thing it gives back. 'input data' = the information you give it. IN SIMPLE WORDS: A prediction is the main answer an AI gives when you give it data. It does not have to be about the future. Like a spam filter saying 'this email is spam'. REAL EXAMPLE: Input a transaction; the prediction is 'fraud, 87% likely'. EXAM TIP: Any main output counts as a prediction, not only forecasts of the future.

New cards

procedural knowledge [3.1.28]

OFFICIAL: knowledge which explicitly indicates the steps to be taken in order to solve a problem or to reach a goal HARD WORDS: 'knowledge' = useful information. 'explicitly' = clearly. 'indicates' = shows. IN SIMPLE WORDS: Procedural knowledge is knowing HOW to do something, step by step. Like a recipe that tells you each step. REAL EXAMPLE: An incident response runbook: isolate the device, reset credentials, notify the client within 24 hours. EXAM TIP: Procedural = HOW. Declarative (3.1.12) = WHAT.

New cards

robot [3.1.29]

OFFICIAL: automation system with actuators that performs intended tasks in the physical world, by means of sensing its environment and a software control system HARD WORDS: 'automation system' = a machine that works by itself. 'actuators' = parts that move or act. 'intended tasks' = the jobs it is meant to do. 'physical world' = real places and things. 'software control system' = programs that tell it what to do. IN SIMPLE WORDS: A robot is a machine with moving parts. It senses what is around it and does jobs in the real world. Software tells it what to do. Like a robot arm in a car factory. REAL EXAMPLE: A warehouse robot that finds a shelf, picks a box and brings it to the packing desk. EXAM TIP: A robot works in the PHYSICAL world. Robotics (3.1.30) is the science of robots.

New cards

robotics [3.1.30]

OFFICIAL: science and practice of designing, manufacturing and applying robots HARD WORDS: 'science and practice' = the study and the real work. 'manufacturing' = making. 'applying' = using. IN SIMPLE WORDS: Robotics is the study and work of making and using robots. Like what a school robot club learns and builds. REAL EXAMPLE: An engineering team developing robots for hospital deliveries works in robotics. EXAM TIP: Robotics = the field. Robot (3.1.29) = the machine.

New cards

semantic computing [3.1.31]

OFFICIAL: field of computing that aims to identify the meanings of computational content and user intentions and to express them in a machine-processable form HARD WORDS: 'semantic' = to do with meaning. 'computational content' = data and information in a computer. 'user intentions' = what a person is trying to do. 'machine-processable form' = a form a computer can use. IN SIMPLE WORDS: This kind of computing tries to understand what information means and what a person wants. Then it stores that meaning so a computer can use it. Like a search box that knows 'cheap flights' means you want to travel. REAL EXAMPLE: A search tool that understands 'laptop stolen from car' means a security incident, not a shopping query. EXAM TIP: Semantic = MEANING. One of the AI concepts in clause 5.6.

New cards

soft computing [3.1.32]

OFFICIAL: field of computing that is tolerant of and exploits imprecision, uncertainty and partial truth to make problem-solving more tractable and robust HARD WORDS: 'tolerant of' = OK with. 'exploits' = makes use of. 'imprecision, uncertainty and partial truth' = things that are fuzzy, unsure or only partly true. 'tractable' = easier to handle. 'robust' = strong and reliable. IN SIMPLE WORDS: Soft computing accepts fuzzy, unsure answers to solve hard problems. It does not insist on perfect answers. Like saying a day is 'quite warm' instead of an exact temperature. REAL EXAMPLE: Fuzzy logic that rates a supplier's risk as 'mostly high' instead of a strict yes or no. EXAM TIP: Includes fuzzy logic and genetic algorithms. Opposite of exact, rule-only computing.

New cards

symbolic AI [3.1.33]

OFFICIAL: AI based on techniques and models that manipulate symbols and structures according to explicitly defined rules to obtain inferences HARD WORDS: 'techniques and models' = methods and tools. 'manipulate symbols and structures' = work with signs and set patterns. 'explicitly defined rules' = rules clearly written by people. 'inferences' = conclusions. IN SIMPLE WORDS: Symbolic AI works by following clear rules written by people. You can read every rule it uses. Like a board game with a printed rule book. REAL EXAMPLE: An expert system with rules like 'IF MFA is missing AND the account is admin THEN risk = high'. EXAM TIP: Symbolic = written RULES, easy to explain. Subsymbolic (3.1.34) = learned from data, harder to explain.

New cards

subsymbolic AI [3.1.34]

OFFICIAL: AI based on techniques and models that use an implicit encoding of information, that can be derived from experience or raw data. HARD WORDS: 'techniques and models' = methods and tools. 'implicit encoding' = storing it in a hidden way, not as readable rules. 'derived from experience or raw data' = learned from examples or data. IN SIMPLE WORDS: Subsymbolic AI learns patterns from data and stores them in a hidden way. You can't read its rules like a list. Like learning to ride a bike by doing it, not from a book. REAL EXAMPLE: A neural network that flags risky emails without anyone being able to point to a written rule it followed. EXAM TIP: Neural networks are subsymbolic. This links to explainability problems (5.15.6).

New cards

★ task [3.1.35]

OFFICIAL: (artificial intelligence) action required to achieve a specific goal HARD WORDS: 'action' = something you do. 'specific goal' = one set target. IN SIMPLE WORDS: A task is a job the AI does to reach a goal. Like 'sort these emails' or 'translate this sentence'. REAL EXAMPLE: Classifying incoming tickets by urgency is a task. Summarising each ticket is another. EXAM TIP: Narrow AI is built for defined tasks; general AI handles a broad range of tasks.

New cards

★ data annotation [3.2.1]

OFFICIAL: process of attaching a set of descriptive information to data without any change to that data HARD WORDS: 'process' = a set of steps. 'attaching' = adding on. 'descriptive information' = notes that describe it. IN SIMPLE WORDS: It means adding labels or notes to data. The data itself does not change. Like putting name tags on photos in an album. REAL EXAMPLE: Tagging 5,000 past emails as 'phishing' or 'safe' so a model can learn from them. EXAM TIP: Annotation adds information but never CHANGES the data. Labels (3.2.10) are one kind of annotation.

New cards

★ data quality checking [3.2.2]

OFFICIAL: process in which data is examined for completeness, bias and other factors which affect its usefulness for an AI system HARD WORDS: 'process' = a set of steps. 'examined' = looked at closely. 'completeness' = nothing missing. 'bias' = unfair leaning one way. 'usefulness' = how helpful it is. IN SIMPLE WORDS: Before an AI uses data, you check it for gaps, unfairness and other problems. Like checking a bag of apples for bad ones before you bake. REAL EXAMPLE: Before training a hiring model, checking whether past applicant data under-represents women or older candidates. EXAM TIP: 42001 Annex A.7.4 asks for data quality requirements. Checking data quality is how you show you meet them.

New cards

★ data augmentation [3.2.3]

OFFICIAL: process of creating synthetic samples by modifying or utilizing the existing data HARD WORDS: 'process' = a set of steps. 'synthetic samples' = new, made-up examples. 'modifying' = changing. 'utilizing' = using. IN SIMPLE WORDS: It means making more practice examples by changing the data you already have. Like flipping a photo so the AI sees more of it. REAL EXAMPLE: Rotating and brightening photos of damaged parts so a defect-detection model has more images to learn from. EXAM TIP: Augmentation creates SYNTHETIC data. It helps when real data is scarce, but it can copy existing bias.

New cards

★ data sampling [3.2.4]

OFFICIAL: process to select a subset of data samples intended to present patterns and trends similar to that of the larger dataset being analysed HARD WORDS: 'process' = a set of steps. 'subset' = a smaller part. 'samples' = single items of data. 'intended to present' = meant to show. 'patterns and trends' = things that repeat or change over time. 'analysed' = studied. IN SIMPLE WORDS: Data sampling means picking a smaller part of your data that still looks like the whole. Like tasting one spoon of soup to judge the whole pot. REAL EXAMPLE: Using 10,000 representative transactions out of 10 million to test a fraud model quickly. EXAM TIP: A bad sample can hide bias. The sample must reflect the patterns of the full dataset.

New cards

★ dataset [3.2.5]

OFFICIAL: collection of data with a shared format HARD WORDS: 'collection' = a group of things kept together. 'shared format' = all set out the same way. IN SIMPLE WORDS: A dataset is a group of data that all follows the same layout. Like a spreadsheet where every row has the same columns. REAL EXAMPLE: A spreadsheet of 2,000 security incidents, each with the same columns: date, type, severity, cause. EXAM TIP: The standard's examples: social media posts from one month, or flower photos all 256x256 pixels.

New cards

exploratory data analysis (EDA) [3.2.6]

OFFICIAL: initial examination of data to determine its salient characteristics and assess its quality HARD WORDS: 'initial examination' = a first look. 'determine' = work out. 'salient characteristics' = the main features that stand out. 'assess its quality' = check how good it is. IN SIMPLE WORDS: It means taking a first look at data. You see what is in it and how good it is. Like flicking through a new book first. REAL EXAMPLE: Charting incident data and noticing half the 'severity' values are blank. EXAM TIP: EDA comes BEFORE training. It often finds missing values and odd data.

New cards

★ ground truth [3.2.7]

OFFICIAL: value of the target variable for a particular item of labelled input data HARD WORDS: 'value' = the answer. 'target variable' = the thing the model is trying to predict. 'labelled input data' = data that already has the right answer attached. IN SIMPLE WORDS: Ground truth is the real, correct answer for a piece of data. You use it to check if the AI got it right. Like the answer sheet for a test. REAL EXAMPLE: An email the security team confirmed was phishing: 'phishing' is its ground truth. EXAM TIP: Ground truth is used to TRAIN and to TEST models. If ground truth is wrong, the model learns the wrong thing.

New cards

imputation [3.2.8]

OFFICIAL: procedure where missing data are replaced by estimated or modelled data HARD WORDS: 'procedure' = a set of steps. 'missing data' = gaps in the data. 'estimated or modelled data' = best guesses worked out from other data. IN SIMPLE WORDS: It means filling a gap in the data with a good guess. Like using the class average to guess a missing height. REAL EXAMPLE: Missing company sizes in a lead list filled with the average for that industry. EXAM TIP: Imputation can add errors or bias. It is a data preparation choice to record under 42001 A.7.6.

New cards

★ input data [3.2.9]

OFFICIAL: data for which an AI system calculates a predicted output or inference HARD WORDS: 'calculates' = works out. 'predicted output' = the answer it gives. 'inference' = a conclusion it reaches. IN SIMPLE WORDS: Input data is the data you give an AI so it can give you an answer. Like the question you type into a search box. REAL EXAMPLE: The text of a new email fed into the phishing detector. EXAM TIP: Input data during live use is called production data (3.2.12).

New cards

★ label [3.2.10]

OFFICIAL: target variable assigned to a sample HARD WORDS: 'target variable' = the thing the model learns to predict. 'assigned' = given. 'sample' = one item of data. IN SIMPLE WORDS: A label is the right answer stuck onto a piece of training data. Like writing 'cat' on a photo of a cat. REAL EXAMPLE: The tag 'urgent' attached to a support ticket in training data. EXAM TIP: Supervised learning needs labels. Unsupervised learning does not.

New cards

★ personally identifiable information (PII; personal data) [3.2.11]

OFFICIAL: any information that (a) can be used to establish a link between the information and the natural person to whom such information relates, or (b) is or can be directly or indirectly linked to a natural person HARD WORDS: 'establish a link' = connect. 'natural person' = a real human being. 'relates' = is about. 'directly or indirectly linked' = connected straight away, or by putting clues together. IN SIMPLE WORDS: Personal information is anything that points to a real person, or could, if you joined some clues together. Like a name, an email, or a photo of a face. REAL EXAMPLE: A name, email address, LinkedIn URL or IP address tied to a person. AI training data containing these brings privacy obligations. EXAM TIP: In Australia, compare 'personal information' under the Privacy Act 1988. PII in training data brings privacy duties.

New cards

★ production data [3.2.12]

OFFICIAL: data acquired during the operation phase of an AI system, for which a deployed AI system calculates a predicted output or inference HARD WORDS: 'acquired' = collected. 'operation phase' = the time when it is used for real. 'deployed' = put into real use. 'calculates' = works out. 'predicted output' = the answer it gives. 'inference' = a conclusion it reaches. IN SIMPLE WORDS: Production data is the real data an AI gets once it is live and people are using it. It is not the practice data it learned from. REAL EXAMPLE: Real customer emails reaching the phishing filter after go-live, as opposed to the training set. EXAM TIP: Production data can differ from training data over time. This is called drift, and it is why AI needs monitoring.

New cards

sample [3.2.13]

OFFICIAL: atomic data element processed in quantities by a machine learning algorithm or an AI system HARD WORDS: 'atomic data element' = the smallest single piece of data. 'processed in quantities' = handled in large amounts. 'algorithm' = a set of steps for solving a problem. IN SIMPLE WORDS: A sample is one single piece of data, like one photo or one email. The AI looks at thousands of them. REAL EXAMPLE: One email in a training set of 50,000 emails. EXAM TIP: A dataset is made of many samples. A label is attached to a sample.

New cards

★ test data (evaluation data) [3.2.14]

OFFICIAL: data used to assess the performance of a final model HARD WORDS: 'assess' = check carefully. 'performance' = how well it works. 'final model' = the finished model. IN SIMPLE WORDS: Test data is data kept aside to check how well the finished model works. The model never learns from it. Like a final exam with questions you haven't seen. REAL EXAMPLE: Keeping 20% of labelled emails aside to measure the finished phishing model's accuracy. EXAM TIP: Training data = learn. Validation data = pick the best model. Test data = final check of the chosen model.

New cards

★ validation data (development data) [3.2.15]

OFFICIAL: data used to compare the performance of different candidate models HARD WORDS: 'performance' = how well it works. 'candidate models' = the different models you are choosing between. IN SIMPLE WORDS: Validation data helps you compare a few models and pick the best one. Like trying on three pairs of shoes before you buy one. REAL EXAMPLE: Trying three versions of a fraud model and choosing the one that scores best on the validation data. Test data is then used once, on the winner. EXAM TIP: Also called development data. Not the same as 'validation' in 3.5.18, which checks fitness for intended use.

New cards

Bayesian network [3.3.1]

OFFICIAL: probabilistic model that uses Bayesian inference for probability computations using a directed acyclic graph HARD WORDS: 'probabilistic model' = a model that works with chances. 'Bayesian inference' = updating a guess as new clues come in. 'probability computations' = working out chances. 'directed acyclic graph' = a diagram of arrows that never loop back. IN SIMPLE WORDS: A Bayesian network is a map of causes and effects that works out how likely things are. Like a map showing how rain makes the road wet and the road makes crashes more likely. REAL EXAMPLE: A model linking 'unpatched server' and 'exposed port' to the probability of a breach. EXAM TIP: One of the example ML methods in clause 5.12.

New cards

decision tree [3.3.2]

OFFICIAL: model for which inference is encoded as paths from the root to a leaf node in a tree structure HARD WORDS: 'inference' = reaching an answer. 'encoded' = stored. 'root' = the starting point. 'leaf node' = an end point. 'tree structure' = a branching diagram. IN SIMPLE WORDS: A decision tree answers by asking yes or no questions, one after another, until it reaches an answer. Like a 'choose your path' game. REAL EXAMPLE: Is the login from overseas? Yes. Is MFA used? No. Then flag as high risk. EXAM TIP: Decision trees are fairly easy to EXPLAIN, which helps with transparency.

New cards

human-machine teaming [3.3.3]

OFFICIAL: integration of human interaction with machine intelligence capabilities HARD WORDS: 'integration' = joining together. 'human interaction' = what people do and say. 'machine intelligence capabilities' = what the AI can do. IN SIMPLE WORDS: Human-machine teaming means people and AI working as one team, each doing what they do best. Like a pilot flying with autopilot. REAL EXAMPLE: AI triages 500 security alerts and a human analyst investigates the top 10 it flags. EXAM TIP: Links to human oversight and controllability (5.15.5).

New cards

★ hyperparameter [3.3.4]

OFFICIAL: characteristic of a machine learning algorithm that affects its learning process HARD WORDS: 'characteristic' = a setting or feature. 'algorithm' = a set of steps for learning. 'process' = the steps it follows. IN SIMPLE WORDS: A hyperparameter is a setting a person chooses before training starts. It changes how the AI learns. Like choosing the oven temperature before you bake. REAL EXAMPLE: Choosing the learning rate or the number of layers before training a model. EXAM TIP: Hyperparameters are set by PEOPLE before training. Parameters (3.3.8) are LEARNED during training.

New cards

★ machine learning (ML) [3.3.5]

OFFICIAL: process of optimizing model parameters through computational techniques, such that the model's behaviour reflects the data or experience HARD WORDS: 'process' = a set of steps. 'optimizing' = making as good as possible. 'model parameters' = the settings inside the model. 'computational techniques' = computer methods. 'reflects' = matches. IN SIMPLE WORDS: The computer tunes its own settings by learning from data or practice. People don't write every rule. Like getting better at throwing by practising. REAL EXAMPLE: A spam filter that learns from millions of labelled emails rather than a list of banned words. EXAM TIP: ML = optimising PARAMETERS from data. It is the most common way to build AI systems today.

New cards

★ machine learning algorithm [3.3.6]

OFFICIAL: algorithm to determine parameters of a machine learning model from data according to given criteria HARD WORDS: 'determine' = work out. 'parameters' = the settings inside the model. 'criteria' = rules for what counts as good. IN SIMPLE WORDS: A machine learning algorithm is the recipe the computer follows to learn from data. Like a study method you follow to learn for a test. REAL EXAMPLE: Gradient descent, decision-tree learning and support vector machines are machine learning algorithms. EXAM TIP: Algorithm = the RECIPE. Model (3.3.7) = what the recipe PRODUCES.

New cards

★ machine learning model [3.3.7]

OFFICIAL: mathematical construct that generates an inference or prediction based on input data or information HARD WORDS: 'mathematical construct' = something built from maths. 'generates' = makes. 'inference' = a conclusion. 'prediction' = the main answer it gives. 'input data' = the information you give it. IN SIMPLE WORDS: A machine learning model is the finished maths that takes in data and gives an answer. Like a trained sniffer dog that smells a bag and says 'yes' or 'no'. REAL EXAMPLE: The trained fraud model file that scores each transaction. EXAM TIP: Model = PRODUCT of training. Algorithm (3.3.6) = the method that made it.

New cards

★ parameter (model parameter) [3.3.8]

OFFICIAL: internal variable of a model that affects how it computes its outputs HARD WORDS: 'internal variable' = a setting inside it. 'computes its outputs' = works out its answers. IN SIMPLE WORDS: A parameter is a setting inside the model that it learns during training. It shapes the answers. Like the knobs on a music mixer. REAL EXAMPLE: The weights inside a neural network. Training adjusts millions of them. EXAM TIP: Parameters are LEARNED. Hyperparameters (3.3.4) are CHOSEN by people.

New cards

★ reinforcement learning (RL) [3.3.9]

OFFICIAL: learning of an optimal sequence of actions to maximize a reward through interaction with an environment HARD WORDS: 'optimal sequence of actions' = the best order of steps. 'maximize a reward' = get as many points as possible. 'interaction with an environment' = trying things and seeing what happens. IN SIMPLE WORDS: Reinforcement learning means learning by trial and error. The AI tries things, gets points for good moves, and slowly learns the best moves. Like training a puppy with treats. REAL EXAMPLE: An AI learning to play a game, or to tune a firewall, by being rewarded when attacks are blocked without disrupting users. EXAM TIP: Three main kinds of ML: supervised, unsupervised, and reinforcement. RL learns from REWARDS, not labels.

New cards

★ retraining [3.3.10]

OFFICIAL: updating a trained model by training with different training data HARD WORDS: 'trained model' = a model that has already learned. 'training data' = data used to teach a model. IN SIMPLE WORDS: It means teaching a model again with new data. Like a footballer learning to play in a new spot on the team. REAL EXAMPLE: Retraining the phishing model each quarter with new attack examples. Under 42001, this is a change that needs control and re-verification. EXAM TIP: Retraining is a CHANGE. Under 42001 it needs control, testing and maybe a fresh risk and impact check.

New cards

★ semi-supervised machine learning [3.3.11]

OFFICIAL: machine learning that makes use of both labelled and unlabelled data during training HARD WORDS: 'labelled data' = data with the right answers attached. 'unlabelled data' = data without answers. 'training' = the learning stage. IN SIMPLE WORDS: The AI learns from a few examples with answers and lots without. Like a teacher marking a few sums, then you do the rest. REAL EXAMPLE: Labelling 1,000 tickets by hand and letting the model also learn from 100,000 unlabelled tickets. EXAM TIP: Between supervised (all labelled) and unsupervised (none labelled).

New cards

★ supervised machine learning [3.3.12]

OFFICIAL: machine learning that makes only use of labelled data during training HARD WORDS: 'labelled data' = data with the right answers attached. 'training' = the learning stage. IN SIMPLE WORDS: The AI learns only from examples that come with the right answers. Like doing a worksheet that has an answer key. REAL EXAMPLE: Training on emails each marked 'phishing' or 'safe'. EXAM TIP: Supervised = ONLY labelled data. Needs ground truth (3.2.7).

New cards

support vector machine (SVM) [3.3.13]

OFFICIAL: machine learning algorithm that finds decision boundaries with maximal margins HARD WORDS: 'algorithm' = a set of steps for learning. 'decision boundaries' = lines that split one group from another. 'maximal margins' = the widest possible gap. IN SIMPLE WORDS: It sorts things into two groups by drawing the widest gap it can between them. Like a referee marking a clear line down the middle. REAL EXAMPLE: Separating 'malware' from 'clean' files using features like file size and permissions. EXAM TIP: One of the example ML methods in clause 5.12.

New cards

★ trained model [3.3.14]

OFFICIAL: result of model training HARD WORDS: 'result' = what you end up with. 'model training' = teaching the model with data. IN SIMPLE WORDS: A trained model is what you get when training is finished. It is ready to use. Like a puppy that has finished obedience school. REAL EXAMPLE: The fraud model ready to deploy after training is complete. EXAM TIP: Trained model = the OUTPUT of training (3.3.15).

New cards

★ training (model training) [3.3.15]

OFFICIAL: process to determine or to improve the parameters of a machine learning model, based on a machine learning algorithm, by using training data HARD WORDS: 'process' = a set of steps. 'determine or improve' = work out or make better. 'parameters' = the settings inside the model. 'algorithm' = the learning method. 'training data' = data used to teach the model. IN SIMPLE WORDS: Training is teaching a model by showing it lots of examples, so it tunes its settings. Like a sports team practising before the real game. REAL EXAMPLE: Running the learning algorithm over 50,000 labelled emails to produce a phishing model. EXAM TIP: Training uses training data. Retraining (3.3.10) repeats it with different data.

New cards

★ training data [3.3.16]

OFFICIAL: data used to train a machine learning model HARD WORDS: 'train' = teach. 'machine learning model' = the maths that learns from data. IN SIMPLE WORDS: Training data is the data a model learns from. If it is unfair or wrong, the model will be too. Like a student who only learns from a textbook full of mistakes. REAL EXAMPLE: The 50,000 labelled emails used to teach the phishing model. If this data is biased, the model will be too. EXAM TIP: Bias in training data becomes bias in the model. 42001 Annex A.7 covers data controls.

New cards

★ unsupervised machine learning [3.3.17]

OFFICIAL: machine learning that makes only use of unlabelled data during training HARD WORDS: 'unlabelled data' = data without answers attached. 'training' = the learning stage. IN SIMPLE WORDS: Unsupervised learning means the AI finds groups and patterns by itself, with no answers given. Like sorting a box of mixed buttons by colour and size without being told how. REAL EXAMPLE: Grouping network traffic into clusters and spotting one unusual cluster that turns out to be an attack. EXAM TIP: Unsupervised = ONLY unlabelled data. Often used for grouping (clustering) and spotting odd items.

New cards

activation function [3.4.1]

OFFICIAL: function applied to the weighted combination of all inputs to a neuron HARD WORDS: 'function' = a maths rule that turns one number into another. 'weighted combination' = adding inputs up, with some counting more than others. 'neuron' = one small working unit in a neural network. IN SIMPLE WORDS: It is the small maths rule inside each neuron. It decides what the neuron passes on, based on what came in. Like a light switch that turns on only when the push is strong enough. REAL EXAMPLE: ReLU, a common activation function, outputs zero for negative values and passes positive values through. EXAM TIP: Activation functions sit inside neurons (3.4.9). ReLU is a common one.

New cards

convolutional neural network (CNN; deep convolutional neural network; DCNN) [3.4.2]

OFFICIAL: feed forward neural network using convolution in at least one of its layers HARD WORDS: 'feed forward neural network' = a network where information flows one way only. 'convolution' = sliding a small filter over data to spot features. 'layers' = rows of neurons. IN SIMPLE WORDS: It is a neural network that slides little filters over data to spot shapes. It is mainly used for pictures. Like a magnifying glass moved across a photo to find edges. REAL EXAMPLE: A CNN that recognises faces in CCTV footage or reads number plates. EXAM TIP: CNNs are the usual choice for images and computer vision (3.7.1).

New cards

convolution [3.4.3]

OFFICIAL: mathematical operation involving a sliding dot product or cross-correlation of the input data HARD WORDS: 'mathematical operation' = a maths step. 'sliding dot product or cross-correlation' = moving a small grid of numbers across data and multiplying as it goes. 'input data' = the data you give it. IN SIMPLE WORDS: It means sliding a small grid across data to find features like edges or lines. Like running your finger over a page to feel raised letters. REAL EXAMPLE: A filter sliding across a photo to find vertical edges. EXAM TIP: Convolution is the key step inside a CNN (3.4.2).

New cards

★ deep learning (deep neural network learning) [3.4.4]

OFFICIAL: (artificial intelligence) approach to creating rich hierarchical representations through the training of neural networks with many hidden layers HARD WORDS: 'approach' = way of doing it. 'rich hierarchical representations' = layers of understanding, from simple to complex. 'training' = teaching with data. 'neural networks' = layers of connected units that learn. 'hidden layers' = middle layers between input and output. IN SIMPLE WORDS: It means using neural networks with many layers. Each layer learns something a bit more complex than the last. Like learning letters, then words, then whole stories. REAL EXAMPLE: Large language models and modern image recognition are built with deep learning. EXAM TIP: 'Deep' = MANY hidden layers. Most modern AI, including chatbots, uses deep learning.

New cards

exploding gradient [3.4.5]

OFFICIAL: phenomenon of backpropagation training in a neural network where large error gradients accumulate and result in very large updates to the weights, making the model unstable HARD WORDS: 'phenomenon' = something that happens. 'backpropagation training' = how a network learns by sending its errors backwards to fix its weights. 'error gradients' = signals showing how wrong it was. 'accumulate' = build up. 'weights' = the strengths of the links. 'unstable' = out of control. IN SIMPLE WORDS: It is a training problem where the network's fixes get bigger and bigger until it goes out of control. Like a microphone squealing when the sound feeds back. REAL EXAMPLE: During training, the loss suddenly jumps to 'infinity' and the model stops learning. EXAM TIP: A known training failure. It makes the model unstable and unreliable.

New cards

feed forward neural network (FFNN) [3.4.6]

OFFICIAL: neural network where information is fed from the input layer to the output layer in one direction only HARD WORDS: 'neural network' = layers of connected units that learn. 'input layer' = where data goes in. 'output layer' = where the answer comes out. IN SIMPLE WORDS: It is a neural network where information only moves forward, from start to finish. It never loops back. Like water flowing down a slide. REAL EXAMPLE: A simple network that takes loan details in and outputs an approval score. EXAM TIP: Feed forward = ONE direction. Recurrent (3.4.10) = has loops.

New cards

long short-term memory (LSTM) [3.4.7]

OFFICIAL: type of recurrent neural network that processes sequential data with a satisfactory performance for both long and short span dependencies HARD WORDS: 'type' = kind. 'recurrent neural network' = a network with loops for handling sequences. 'processes' = handles. 'sequential data' = data in order, like words in a sentence. 'long and short span dependencies' = links between things close together and far apart. IN SIMPLE WORDS: It is a kind of neural network that remembers things from both a moment ago and much earlier. Like remembering the start of a story when you reach the end. REAL EXAMPLE: Predicting the next word in a sentence, or spotting unusual patterns in a long log file. EXAM TIP: LSTM is a special kind of RNN (3.4.10) that copes better with long sequences.

New cards

★ neural network (NN; neural net; artificial neural network) [3.4.8]

OFFICIAL: (artificial intelligence) network of one or more layers of neurons connected by weighted links with adjustable weights, which takes input data and produces an output HARD WORDS: 'network' = linked parts. 'layers of neurons' = rows of small working units. 'weighted links' = connections with a strength. 'adjustable weights' = strengths that can change as it learns. 'input data' = the data you give it. IN SIMPLE WORDS: It is rows of small linked units. Each link has a strength that changes as it learns. Data goes in one end and an answer comes out the other. Like a team passing a ball along. REAL EXAMPLE: The network inside an image classifier that takes pixels in and outputs 'cat' or 'dog'. EXAM TIP: Neural networks are subsymbolic AI and power deep learning. Their inner workings are hard to explain.

New cards

neuron [3.4.9]

OFFICIAL: (artificial intelligence) primitive processing element which takes one or more input values and produces an output value by combining the input values and applying an activation function on the result HARD WORDS: 'primitive processing element' = a very simple working part. 'input values' = numbers that come in. 'output value' = the number it sends out. 'activation function' = the small maths rule that decides what to pass on. IN SIMPLE WORDS: A neuron is the tiny basic part of a neural network. It adds up what comes in and decides what to send on. Like one person in a relay team. REAL EXAMPLE: One node that adds up weighted signals from the previous layer and passes the result on. EXAM TIP: Neuron = the unit. Activation function (3.4.1) = the rule inside it.

New cards

recurrent neural network (RNN) [3.4.10]

OFFICIAL: neural network in which outputs from both the previous layer and the previous processing step are fed into the current layer HARD WORDS: 'neural network' = layers of connected units that learn. 'previous layer' = the row before. 'previous processing step' = the step before in time. 'fed into' = passed to. IN SIMPLE WORDS: It is a neural network with a loop. What it worked out a moment ago feeds back in, so it can follow things in order. Like reading a sentence word by word. REAL EXAMPLE: Processing a sentence word by word, remembering earlier words. EXAM TIP: Recurrent = has a LOOP for sequences. LSTM (3.4.7) is an improved RNN.

New cards

★ accountable [3.5.1]

OFFICIAL: answerable for actions, decisions and performance HARD WORDS: 'answerable' = must explain and take the blame or credit. 'performance' = how well things went. IN SIMPLE WORDS: Being accountable means you must answer for what you did, what you decided and how it turned out. Like the captain who answers to the coach after a game. REAL EXAMPLE: The Head of Data is accountable for the hiring AI's decisions and must answer to the board if it discriminates. EXAM TIP: Accountable = answers FOR results. The AI system itself can't be accountable; people and organisations are.

New cards

★ accountability [3.5.2]

OFFICIAL: state of being accountable HARD WORDS: 'state' = condition. 'accountable' = must answer for actions, decisions and results. IN SIMPLE WORDS: It means being the one who must answer for something. It usually comes from a job you were given by law, contract or your boss. REAL EXAMPLE: A 42001 role description making the AI Product Owner accountable for model risk sign-off. Accountability can't be handed to the AI itself. EXAM TIP: Comes from ALLOCATED responsibility (Note 1). 42001 clause 5.3 and A.3.2 assign it.

New cards

★ availability [3.5.3]

OFFICIAL: property of being accessible and usable on demand by an authorized entity HARD WORDS: 'property' = a quality it has. 'accessible' = you can get to it. 'usable on demand' = works when you need it. 'authorized entity' = a person or system that is allowed. IN SIMPLE WORDS: Available means the right people can get to it and use it when they need it. Like a lift that works whenever you press the button. REAL EXAMPLE: The fraud model's scoring service is available 99.9% of the time during business hours. EXAM TIP: Availability is the 'A' in the CIA triad (confidentiality, integrity, availability).

New cards

★ bias [3.5.4]

OFFICIAL: systematic difference in treatment of certain objects, people or groups in comparison to others HARD WORDS: 'systematic difference' = a difference that happens again and again, not by chance. 'treatment' = how something is handled. 'in comparison to others' = compared with other groups. IN SIMPLE WORDS: Bias is when some people or things are treated differently every time, in a pattern. Like a game where one team always gets the easier side of the field. REAL EXAMPLE: A CV-screening model consistently scores women lower for the same experience: unwanted bias. EXAM TIP: The definition is NEUTRAL. Some bias is wanted; UNWANTED bias causes unfairness (5.15.9).

New cards

★ control [3.5.5]

OFFICIAL: purposeful action on or in a process to meet specified objectives HARD WORDS: 'purposeful action' = something done on purpose. 'process' = a set of steps. 'specified objectives' = set goals. IN SIMPLE WORDS: It means doing something on purpose to keep a job on track and meet its goals. Like a coach changing players to win the game. REAL EXAMPLE: Requiring human review of every AI-rejected loan application is a control. EXAM TIP: 22989 'control' = action on a process. 42001's 'control' (3.21) = a measure that keeps or changes RISK.

New cards

★ controllability (controllable) [3.5.6]

OFFICIAL: property of an AI system that allows a human or another external agent to intervene in the system’s functioning HARD WORDS: 'property' = a quality it has. 'external agent' = a person or system outside it. 'intervene' = step in. 'functioning' = how it works. IN SIMPLE WORDS: It means a person or another system can step in and change or stop what the AI is doing. Like a big red stop button on a machine. REAL EXAMPLE: An operator can pause the AI chatbot or override its answer at any time. 42001 expects this kind of human oversight. EXAM TIP: Controllability enables HUMAN OVERSIGHT. Linked to heteronomy (3.1.16).

New cards

★ explainability [3.5.7]

OFFICIAL: property of an AI system to express important factors influencing the AI system results in a way that humans can understand HARD WORDS: 'property' = a quality it has. 'express' = show. 'factors influencing' = things that affect. IN SIMPLE WORDS: It means the AI can show the main reasons behind its answer, in a way people understand. It answers 'why?'. Like a teacher showing how they marked your test. REAL EXAMPLE: The loan model shows: declined mainly because of income and recent missed payments. EXAM TIP: Explainability answers WHY. It does not argue the answer was the best one (Note 1).

New cards

★ predictability [3.5.8]

OFFICIAL: property of an AI system that enables reliable assumptions by stakeholders about the output HARD WORDS: 'property' = a quality it has. 'enables reliable assumptions' = lets people safely guess. 'stakeholders' = anyone who can affect or be affected. 'output' = the answer it gives. IN SIMPLE WORDS: It means people can tell ahead of time what kind of answer the AI will give. Like knowing a vending machine gives a drink when you press the button. REAL EXAMPLE: Staff know the classifier always routes invoices over $10,000 to a human, so its behaviour is predictable. EXAM TIP: Predictability = people can FORESEE outputs. Different from reliability (consistent results).

New cards

★ reliability [3.5.9]

OFFICIAL: property of consistent intended behaviour and results HARD WORDS: 'property' = a quality it has. 'consistent' = the same each time. 'intended behaviour' = acting the way it was meant to. IN SIMPLE WORDS: It means it behaves the way it should and gives the right results, every time. Like a watch that always shows the right time. REAL EXAMPLE: The model gives the same correct answer for the same type of input, day after day. Reliability = consistent; robustness = copes with anything; resilience = bounces back. EXAM TIP: Reliability = CONSISTENT. Robustness = copes with ANYTHING. Resilience = BOUNCES BACK.

New cards

★ resilience [3.5.10]

OFFICIAL: ability of a system to recover operational condition quickly following an incident HARD WORDS: 'ability' = being able to. 'recover operational condition' = get back to working normally. 'incident' = something going wrong. IN SIMPLE WORDS: It means bouncing back fast after something goes wrong. Like a phone that restarts and works again after it freezes. REAL EXAMPLE: After a cloud outage, the AI service fails over and is running again within five minutes. EXAM TIP: Resilience = RECOVERY after an incident. Robustness = not failing in the first place.

New cards

★ risk [3.5.11]

OFFICIAL: effect of uncertainty on objectives HARD WORDS: 'effect' = what happens because of something. 'uncertainty' = not knowing for sure. 'objectives' = goals. IN SIMPLE WORDS: Risk is how not knowing what will happen could change whether you reach your goals. It can be good or bad. Like rain that might spoil a picnic. REAL EXAMPLE: Uncertainty about whether the training data is biased threatens the objective of fair hiring decisions. EXAM TIP: ISO 31000 definition. Note it says 'on OBJECTIVES' here; 42001's version (3.7) is shorter.

New cards

★ robustness [3.5.12]

OFFICIAL: ability of a system to maintain its level of performance under any circumstances HARD WORDS: 'ability' = being able to. 'maintain' = keep. 'level of performance' = how well it works. 'under any circumstances' = no matter what happens. IN SIMPLE WORDS: It means it keeps working well no matter what is thrown at it, even odd or tricky inputs. Like a phone that still works after you drop it. REAL EXAMPLE: The image classifier still works on blurry photos, odd lighting or images deliberately tampered with. EXAM TIP: Robustness = holds up against odd, noisy or attacked inputs (5.15.2).

New cards

★ stakeholder [3.5.13]

OFFICIAL: any individual, group, or organization that can affect, be affected by or perceive itself to be affected by a decision or activity HARD WORDS: 'individual' = one person. 'affect' = change something. 'perceive itself to be affected' = believe it is touched by it. IN SIMPLE WORDS: It means anyone who can change a decision, is changed by it, or feels changed by it. Like everyone with an opinion on a new school uniform. REAL EXAMPLE: For a hiring AI: applicants, recruiters, HR, the board, regulators and unions are all stakeholders. EXAM TIP: Same idea as 'interested party' in 42001 (3.2).

New cards

★ transparency, organization [3.5.14]

OFFICIAL: (organization) property of an organization that appropriate activities and decisions are communicated to relevant stakeholders in a comprehensive, accessible and understandable manner HARD WORDS: 'property' = a quality it has. 'appropriate' = suitable. 'communicated' = told. 'relevant stakeholders' = the people who care. 'comprehensive, accessible and understandable' = complete, easy to get to and easy to follow. IN SIMPLE WORDS: It means the company tells people clearly what it is doing and deciding. Like a school sending parents a clear newsletter. REAL EXAMPLE: A company publishes how and where it uses AI in its hiring process. (Compare 3.5.15, which is about the system.) EXAM TIP: Two transparency terms: this one is about the ORGANISATION. 3.5.15 is about the SYSTEM.

New cards

★ transparency, system [3.5.15]

OFFICIAL: (system) property of a system that appropriate information about the system is made available to relevant stakeholders HARD WORDS: 'property' = a quality it has. 'appropriate information' = the right facts. 'made available' = shared. 'relevant stakeholders' = the people who care. IN SIMPLE WORDS: It means the right facts about the AI itself are shared with the people who need them. Like the ingredients label on food. REAL EXAMPLE: A model card stating the chatbot's purpose, data sources, known limits and accuracy. (Compare 3.5.14, which is about the organisation.) EXAM TIP: System transparency covers features, performance, limits, data sources and design choices. Some details can stay private for security.

New cards

★ trustworthiness [3.5.16]

OFFICIAL: ability to meet stakeholder expectations in a verifiable way HARD WORDS: 'ability' = being able to. 'stakeholder expectations' = what people hope for and need. 'verifiable' = can be checked and proven. IN SIMPLE WORDS: It means doing what people expect in a way you can prove. Like a builder showing safety certificates, not just saying 'trust me'. REAL EXAMPLE: Not 'trust us, it's fair', but 'here is the bias test showing it's fair'. The word 'verifiable' is the key. EXAM TIP: Key word: VERIFIABLE. Covers reliability, security, privacy, safety, transparency and more (Note 2).

New cards

★ verification [3.5.17]

OFFICIAL: confirmation, through the provision of objective evidence, that specified requirements have been fulfilled HARD WORDS: 'confirmation' = proof that it is so. 'provision of objective evidence' = giving fair, real proof. 'specified requirements' = the rules written down for it. 'fulfilled' = met. IN SIMPLE WORDS: It means proving, with real evidence, that it meets the rules written for it. 'Did we build it right?' Like checking a cake matches the recipe. REAL EXAMPLE: Testing confirms the model meets the spec: accuracy at least 95% and response under 2 seconds. EXAM TIP: Verification = matches the SPEC. Validation (3.5.18) = right for its real use.

New cards

★ validation [3.5.18]

OFFICIAL: confirmation, through the provision of objective evidence, that the requirements for a specific intended use or application have been fulfilled HARD WORDS: 'confirmation' = proof that it is so. 'provision of objective evidence' = giving fair, real proof. 'requirements' = what it must do. 'specific intended use or application' = the real job it is meant for. 'fulfilled' = met. IN SIMPLE WORDS: It means proving, with real evidence, that it actually works for the job it is meant for. 'Did we build the right thing?' Like checking the cake tastes good to the guests. REAL EXAMPLE: A trial with real recruiters shows the screening tool actually helps them pick good candidates fairly. It can pass verification and still fail validation. EXAM TIP: A system can pass verification and still fail validation. Exams love this pair.

New cards

automatic summarization [3.6.1]

OFFICIAL: task of shortening a portion of natural language content or text while retaining important semantic information HARD WORDS: 'task' = a job. 'natural language' = everyday human language. 'retaining' = keeping. 'semantic information' = the meaning. IN SIMPLE WORDS: It means making text shorter but keeping the main points. Like turning a long chapter into a few lines on a revision card. REAL EXAMPLE: Turning a 20-page audit report into a 5-line summary. EXAM TIP: One of the NLP tasks listed in clause 9.2.

New cards

dialogue management [3.6.2]

OFFICIAL: task of choosing the appropriate next move in a dialogue based on user input, the dialogue history and other contextual knowledge, to meet a desired goal HARD WORDS: 'task' = a job. 'appropriate next move' = the right thing to say or do next. 'dialogue' = a conversation. 'contextual knowledge' = what it knows about the situation. 'desired goal' = what it is trying to reach. IN SIMPLE WORDS: It means a chat system picking the right thing to say next, based on what was said before. Like a shop helper asking for your size before fetching shoes. REAL EXAMPLE: A chatbot deciding to ask for an order number before answering a delivery question. EXAM TIP: This is the 'brain' of a chatbot's conversation flow.

New cards

emotion recognition [3.6.3]

OFFICIAL: task of computationally identifying and categorizing emotions expressed in a piece of text, speech, video or image or combination thereof HARD WORDS: 'task' = a job. 'computationally identifying' = a computer finding. 'categorizing' = sorting into groups. 'combination thereof' = a mix of these. IN SIMPLE WORDS: It means a computer spotting feelings like anger or joy in words, voices, faces or videos. Like a friend who can tell you're upset from your voice. REAL EXAMPLE: Flagging an angry customer call so a supervisor can step in. EXAM TIP: High-risk use: emotion recognition at work or school raises privacy and fairness worries.

New cards

information retrieval (IR) [3.6.4]

OFFICIAL: task of retrieving relevant documents or parts of documents from a dataset, typically based on keyword or natural language queries HARD WORDS: 'task' = a job. 'retrieving' = finding and fetching. 'relevant' = useful for what you asked. 'dataset' = a collection of data. 'queries' = questions or search words. IN SIMPLE WORDS: It means finding the right documents, or parts of them, from a big pile when you ask. Like a librarian finding the right book for you. REAL EXAMPLE: Searching a policy library for 'password rotation' and getting the right section back. EXAM TIP: IR finds DOCUMENTS. Question answering (3.6.14) gives the ANSWER itself.

100

New cards

machine translation (MT) [3.6.5]

OFFICIAL: task of automated translation of text or speech from one natural language to another using a computer system HARD WORDS: 'task' = a job. 'automated translation' = a computer changing one language into another. 'natural language' = everyday human language. IN SIMPLE WORDS: It means a computer turning words from one language into another. Like an app that turns Hindi into English. REAL EXAMPLE: Translating a Singapore client's Mandarin policy into English. EXAM TIP: Works on text OR speech.