A plain-English primer for anyone new to AI: what AI is, how it learns, why it goes wrong, and why it needs rules. Read this first, then start Domain 1.
Normal software follows a recipe a person wrote step by step. AI is software that learned its own recipe by studying lots of examples, so it can handle things nobody wrote a rule for, like spotting a cat in a photo it has never seen.
diagram
Regular software runs rules a person wrote. AI learns a pattern from examples and uses it to make a good guess about something new.
Key points
Regular software: a person writes every rule, so the same input always gets the same answer.
AI: the system works out its own rules from examples, then uses them on new things it has never seen.
The official definitions from the OECD and the EU AI Act hinge on one word, infer: the system works out how to turn inputs into outputs.
Outputs can be predictions, content, recommendations or decisions, and they can change things in the real or digital world.
A model is the learned recipe; an AI system is the whole package around it, including data, software and the people using it.
Exam tip
If an answer describes fixed if-then rules written entirely by people, it is not AI under the OECD or EU definitions. Look for the word infer.
Training an AI is like a student studying with flashcards. It looks at an example, guesses, checks the back of the card and adjusts. After millions of cards it gets good at guessing, but only as good as the cards it studied.
diagram
The flashcard loop: guess, check, adjust, repeat millions of times. When the loop stops, the learned settings are saved as the model.
Key points
Training data is the stack of flashcards: the examples the model learns from.
Each wrong guess nudges the model's internal settings so the next guess is a little better.
The finished model is a file of learned settings; it no longer needs the flashcards to answer.
Bad or lopsided flashcards teach bad or lopsided habits, which is why data governance matters so much.
Testing on cards the model never studied shows whether it really learned or just memorized.
Exam tip
Many governance controls trace back to data: where it came from, whether you may use it, and whether it represents the people the system will affect.
Learning to ride a bike takes weeks; riding to the store takes minutes. AI works the same way: training is the slow, expensive learning part, and using the model, called inference, is the quick everyday part.
diagram
Training is the slow learning phase that makes the model. Inference is every quick use after that. Monitoring watches for drift and sends the model back for retraining.
Key points
Training happens once or now and then, uses lots of data and computing power, and produces the model.
Inference is each time someone uses the model: a question goes in and an answer comes out in seconds.
Fine-tuning is extra training on a smaller, focused set of examples to teach a model a specific job.
Models drift: the world changes after training, so answers slowly get worse unless someone keeps checking.
Governance applies at both stages: what goes into training, and how the model behaves when people use it.
Exam tip
When a question describes results getting worse over time in production, think drift and monitoring, not a bug in the original code.
A large language model is a very well-read autocomplete. It read a huge amount of text and learned which words tend to come next, so it writes an answer one small piece at a time. It sounds just as sure of itself when it is guessing.
diagram
A large language model writes one token at a time. Each new token is added to the text and the model predicts the next one, until the answer is done.
Key points
A large language model (LLM) predicts the next small chunk of text, called a token, over and over to build an answer.
Because it learned from huge amounts of text, it can write, summarize, translate and answer questions.
Chat assistants are built on a large language model plus extra training to follow instructions and stay safe.
It does not look facts up unless it is connected to a search tool or document store; otherwise it writes from what it learned.
A general-purpose AI (GPAI) model can do many different jobs, and the EU AI Act has its own rules for the companies that make them.
Exam tip
Expect questions to treat the large language model as one component. Governance looks at the whole system built around it: data, prompts, tools and people.
Predictive AI answers what is likely, like a weather forecast or a spam filter. Generative AI makes something new, like an email or a picture. Both learn from examples, but they go wrong in different ways.
Predictive AI
Spam filters
Loan and fraud scores
Weather and demand forecasts
Both
Learn from examples
Need good data
Need testing and monitoring
Generative AI
Writes emails and code
Makes images and audio
Can invent facts
Both kinds learn from examples. Predictive AI judges what is likely; generative AI makes something new. Agents built on generative AI can also take actions.
Key points
Predictive AI scores, sorts or forecasts: fraud flags, loan scores, spam filters and demand forecasts.
Generative AI creates new text, images, audio, video or code from a prompt.
Predictive AI that scores people, for example for jobs or credit, can fall into the EU AI Act high-risk group.
Generative AI can invent facts that sound real, often called hallucinations, and can produce harmful content.
Agentic AI goes a step further: it plans and takes actions with tools, like sending emails or changing records.
Exam tip
Match the risk to the type: unfair scoring points to predictive AI and bias testing, while invented facts and unsafe content point to generative AI and output controls.
AI makes mistakes for simple reasons: it learned from lopsided examples, it is guessing about something new, or the world changed since it studied. The trouble is that it often sounds just as confident when it is wrong.
diagram
Five everyday reasons AI goes wrong. Each one has its own fix, which is why governance uses several controls instead of one.
Key points
Bias: if the examples left some groups out, the model can treat those groups unfairly.
Hallucination: a generative model can make up facts, quotes or sources that sound real.
Drift: the world changes after training, so yesterday's patterns stop matching today.
Brittleness: a small, odd change to an input can flip the answer in a way a person never would.
Opacity: many models cannot show their reasoning in plain terms, which makes errors harder to spot and explain.
Exam tip
Find the root cause in the question. Unfair outcomes point to data and fairness testing, invented facts to output checks and human review, and slow decline to monitoring.
Attackers do not always break in; sometimes they just talk the AI into misbehaving. Picture someone slipping a fake instruction into a pile of homework the teacher is grading. AI security is about spotting and blocking those tricks.
diagram
Indirect prompt injection: the attacker never talks to the AI directly. The hidden instruction rides in on content the AI was asked to read.
Key points
Prompt injection: sneaky instructions make the AI ignore its rules; it is LLM01 in the OWASP 2026 list.
Data poisoning: bad examples slipped into training data teach the model the wrong lesson on purpose.
Sensitive information disclosure: the AI leaks private data it saw or was given; it is LLM02 in the OWASP 2026 list.
Model theft: someone copies a model by stealing its files or by querying it many times to rebuild it.
MITRE ATLAS is a public catalog of real attacks on AI systems, built like MITRE ATT&CK but for AI.
Exam tip
If the attack arrives through something the AI reads, such as a web page, email or file, the answer is usually indirect prompt injection.
A chatbot only talks. An AI agent can also do things, like book a meeting, send an email or change a record, using tools it has been given. That is like handing a new intern the office keys, so you choose carefully which doors they can open.
diagram
An agent breaks a goal into steps and calls tools. Low-risk steps run on their own; a risky step waits for a person to approve it.
Key points
An agent plans steps toward a goal and calls tools, such as email, search or a database, to carry them out.
The more tools and access an agent has, the bigger the damage if it is tricked or makes a mistake.
Excessive agency, giving an AI more power than the job needs, is LLM03 in the OWASP 2026 list.
Good practice is least privilege, a person approving risky actions, and a log of everything the agent did.
The OWASP Top 10 for Agentic Applications 2026 runs from ASI01 Agent Goal Hijack to ASI10 Rogue Agents.
Exam tip
For agent questions, the safest answer limits what the agent can touch and keeps a person in the loop for high-impact actions.
Cars are useful and dangerous, so we have driver's licenses, crash tests, road rules, registration and police. AI governance is the same idea for AI: who may use it, how it is built safely, how it is checked, and who answers for it.
diagram
Everything we do to keep cars safe has an AI governance twin. The exam is mostly about the right-hand side.
Key points
Driver's license: AI literacy training so people know how to use AI safely.
Crash tests before sale: testing and evaluation before an AI system goes live.
Road rules: laws like the EU AI Act plus the organization's own AI policy.
Registration: an inventory of every AI system with a named owner.
Police and inspections: monitoring, incident response and audits once the AI is in use.
Exam tip
Governance is never one tool. Answers that combine policy, people, testing and monitoring usually beat answers that rely on a single technical control.
Rules for AI come in layers. Laws are the must-do rules from governments. Frameworks are helpful playbooks. Standards are detailed checklists you can be certified against. Company policy turns all of it into how we do it here.
diagram
Four layers of rules. Laws say what you must do, frameworks and standards say how to do it well, and company policy makes it real day to day.
Key points
Laws: the EU AI Act is the big one; after the July 2026 Digital Omnibus, Annex III high-risk duties apply from 2 December 2027.
Frameworks: the NIST AI Risk Management Framework (AI RMF) has four functions: Govern, Map, Measure and Manage.
Standards: ISO/IEC 42001 sets up an AI management system that an organization can be certified against.
Principles: the OECD AI Principles are shared values many countries signed up to, such as transparency and accountability.
Company policy: an AI policy, acceptable use rules and an AI governance council put it all into practice.
Exam tip
Know which layer a question is about: the EU AI Act is mandatory law, the NIST AI RMF is voluntary guidance, and ISO/IEC 42001 is the certifiable standard.
The exam follows the life of an AI project like a story: understand what AI is and what is fair, decide the plan and the rules, find what could go wrong, build it safely, then keep watching it and check it with audits.
diagram
The five exam domains follow one AI project from idea to audit, then loop back as lessons learned change the plan.
Key points
Domain 1, AI basics and ethics (17%): what AI is and what fair, responsible AI looks like.
Domain 2, strategy and governance (24%): goals, policies, frameworks and laws such as the EU AI Act.
Domain 3, risk management (16%): finding, scoring and treating AI risks, including risks from suppliers.
Domain 4, development and implementation (27%): building AI securely and protecting privacy.
Domain 5, operations and audit (16%): incidents, continuity, testing and audits once the AI is running.
Exam tip
Domains 2 and 4 make up just over half the exam, so give extra time to governance frameworks, the EU AI Act and secure AI design.