The EU AI Act (Regulation (EU) 2024/1689) is the world's first comprehensive legal framework for artificial intelligence, and it is the most consequential piece of legislation you will encounter on the AIGP exam. Yet most candidates approach it as a list of definitions to memorise. That is the wrong frame. The Act is a risk architecture — a structured hierarchy of harm, obligation, and enforcement that you must be able to navigate quickly under exam pressure.
This guide maps every tier of the risk pyramid to its real-world regulatory impact, enforcement dates, and the specific exam traps that separate passing candidates from failing ones. All content is aligned to the 2025/2026 AIGP Body of Knowledge (v2.0.1) and reflects the "AI Omnibus" updates that expanded the list of prohibited practices in force since February 2025.
The Risk-Based Architecture: Why It Matters for the Exam
The Act's foundational principle is proportionality: the more dangerous the AI system, the heavier the compliance burden. Risk is not determined by the underlying technology or the sophistication of the model. It is determined by the intended purpose and deployment context. A chatbot is a limited-risk system when used for customer service; it becomes a high-risk system the moment it is deployed to triage medical patients. This context-dependence is the single most tested concept across all AIGP exam forms.
The Act also has extraterritorial reach. It applies to any provider, deployer, importer, or distributor whose AI system's outputs affect persons within the European Union — regardless of where that organisation is headquartered. A company based in Singapore that uses a credit-scoring model affecting EU residents is fully within scope.
Tier 1 — Unacceptable Risk
Prohibited Practices
Tier 2 — High Risk
Mandatory Conformity Obligations
Tier 3 — Limited Risk
Transparency Obligations (Article 50)
Tier 4 — Minimal / No Risk
No Mandatory Obligations
e.g. spam filters,
video games
Tier 1: Unacceptable Risk — The Nine Prohibitions
Prohibited practices represent AI applications the EU legislature determined pose such a fundamental threat to human dignity and fundamental rights that no compliance framework can adequately mitigate them. The "AI Omnibus" update expanded the original list to nine prohibited categories, all enforceable since February 2025.
The exam tests candidates' ability to identify the boundary conditions — the narrow exceptions that make a prohibited practice legal in very specific circumstances. Memorising the prohibitions is necessary but insufficient. You must know where the exceptions sit.
Prohibited Practice
Key Exam Nuance / Exception
Social Scoring
Banned for both governments and private entities when it causes unfavourable treatment in unrelated social contexts. Not all scoring systems are prohibited — only those with cross-context harm.
Emotion Recognition
Prohibited in workplace and educational settings. Narrow exception for medical or safety purposes (e.g. monitoring pilot or driver fatigue). This exception is a frequent exam trap.
Exploitative AI
AI that exploits vulnerabilities (age, disability, socioeconomic status) to materially distort behaviour and cause harm. Both elements — exploitation and harm — must be present.
Subliminal Manipulation
AI using techniques below the threshold of conscious awareness to influence behaviour in ways that cause physical or psychological harm. The harm criterion is essential.
Biometric Categorisation
Prohibited when used to deduce protected characteristics (race, religion, sexual orientation). Not all biometric categorisation is banned — only inference of sensitive attributes.
Predictive Policing
Profiling individuals to predict criminal risk based solely on traits or characteristics. Systems using actual behavioural evidence are outside this prohibition.
Untargeted Scraping
Indiscriminate scraping of facial images from the internet or CCTV to build recognition databases. The word "untargeted" is the operative qualifier in this prohibition.
Real-Time Biometric ID
Prohibited for law enforcement in public spaces, except for specific authorised threats: searching for missing children, preventing imminent terrorist attacks, or identifying suspects in serious crimes.
"Nudification" Apps OMNIBUS
AI that generates non-consensual sexually explicit or intimate imagery. This category was added by the Omnibus update and is a likely new exam topic for 2025/2026 forms.
⚠ High-Probability Exam Trap: The Article 25 Provider Shift
A Deployer becomes a Provider — and assumes all provider obligations — in three specific scenarios: (1) they place their own name or trademark on a high-risk AI system already on the market, (2) they make a substantial modification to a high-risk system that keeps it high-risk, or (3) they modify the intended purpose of a non-high-risk system so that it becomes high-risk. Exam questions frequently present a scenario where a company "rebrands" a vendor's model — the correct answer is that they have become the provider.
Tier 2: High-Risk AI — The Compliance Core
High-risk AI represents the regulatory centre of gravity in the Act. The compliance obligations here — risk management systems, data governance, technical documentation, human oversight — are the subject of the majority of AIGP exam questions. A system reaches high-risk status through one of two channels.
Channel 1 — Annex I Products: The AI system is used as a safety component in a product already governed by existing EU safety legislation, such as medical devices, machinery, or toys. The risk classification flows from the product category.
Channel 2 — Annex III Use Cases: The system falls within one of eight specific deployment contexts that the legislature identified as carrying a significant risk of harm or discrimination, regardless of the underlying technology.
The Eight High-Risk Annex III Use Cases
1
Biometrics
Remote identification and categorisation not already prohibited. Justification: high risk of mass surveillance and algorithmic bias at scale.
2
Critical Infrastructure
Safety components in energy, water, or transport networks. Justification: failure could jeopardise lives at a population scale.
3
Education
Admissions decisions and academic performance evaluation. Justification: biased scoring can permanently alter career trajectories. Real case: the UK A-Level algorithm.
4
Employment
CV sorting, promotion decisions, and task allocation. Justification: impacts earning ability and labour rights. Real case: Amazon's recruitment tool penalising women.
5
Essential Services
Credit scoring, insurance pricing, welfare benefits. Justification: potential for systemic financial exclusion. Real cases: Apple Card gender disparity, Dutch Childcare Benefit Scandal.
6
Law Enforcement
Reliability of evidence and recidivism risk assessment. Justification: power imbalances in criminal proceedings can lead to irreversible fundamental rights violations.
7
Migration & Border Control
Security risk assessments and visa examinations. Justification: decisions affect individuals in highly vulnerable situations with limited legal recourse.
8
Administration of Justice
AI tools used by courts to research or interpret law. Justification: direct impact on rule of law and judicial independence.
Article 6(3) — The Narrow Non-High-Risk Exception
A system otherwise falling within an Annex III area is not classified as high-risk if it performs a narrow procedural task (e.g. converting unstructured data to structured format), merely improves the tone of a document, or detects patterns without overriding human review. Critical exam rule: this exception never applies if the system is used for profiling. Candidates frequently overlook this absolute carve-out and select the exception when a profiling element is present.
Mandatory Obligations for High-Risk Providers
For every high-risk system that clears Article 6(3), the provider must implement and maintain a comprehensive compliance stack before placing the system on the EU market. The exam frequently asks which obligation applies in a specific fact pattern — understanding the purpose behind each requirement is more reliable than rote memorisation.
Obligation Category
What It Requires
Risk Management
An iterative system to identify, analyse, and mitigate risks throughout the entire AI lifecycle — not a one-time pre-launch audit.
Data Governance
Training, validation, and test datasets must be relevant, representative, and "to the best extent possible" free from errors. Note the qualified standard — perfection is not mandated.
Technical Documentation
Must be prepared before placing the system on the market. SMEs and small mid-cap companies benefit from simplified documentation requirements.
Record-Keeping
Automatic event logging for traceability. Logs must be retained for a minimum of 6 months. This specific figure has appeared directly in exam questions.
Transparency
Provide clear "Instructions for Use" to deployers, explicitly detailing system limitations and potential risks.
Human Oversight
Systems must be designed with override and halt capabilities. The goal is to prevent automation bias — the tendency of operators to defer to AI outputs uncritically.
Accuracy & Robustness
Maintain documented performance benchmarks and resilience against cybersecurity threats, including data poisoning and adversarial attacks.
Four Case Studies That Defined the Risk Tiers
The AIGP exam is case-study-driven. The EU AI Act's high-risk categories were not invented by legal theorists — they emerged directly from documented failures that caused real harm to real people. Knowing the case that corresponds to each sector allows you to answer scenario questions with precision rather than guesswork.
Essential Public Services
The Dutch Childcare Benefit Scandal
A fraud-detection algorithm wrongly accused approximately 20,000 families — disproportionately from ethnic minorities — based on biased risk indicators, causing financial ruin and family separations. This real-world failure directly shaped the "Essential Public Services" high-risk category and established the legal foundation for mandatory bias audits in welfare systems.
Employment
Amazon's Recruitment Tool
A machine learning CV-review system trained on a decade of male-dominated hiring decisions learned to systematically penalise female candidates by downgrading CVs containing the word "women's." The system was scrapped internally but its discovery became the definitive justification for classifying recruitment AI as high-risk under Annex III.
Essential Private Services
Apple Card Credit Limits
A credit-scoring model assigned women systematically lower credit limits than men with identical financial profiles by using household spending patterns as a proxy for gender. No protected characteristic was explicitly used — the discrimination emerged from facially neutral variables. This case illustrates why surface-level bias audits are insufficient and why proxy discrimination is a core AIGP exam concept.
Education
UK A-Level Algorithm (2020)
An algorithm designed to standardise lockdown-era exam results penalised high-achieving students at historically lower-performing state schools by anchoring grades to historical institutional averages rather than individual performance. Students from disadvantaged backgrounds were disproportionately downgraded, defining the "Education" high-risk area and establishing the case for mandatory human review overrides.
Tier 3: Limited Risk — Transparency Under Article 50
Limited-risk systems are not regulated out of existence — they are regulated into honesty. The obligation is disclosure, not conformity assessment. Article 50 requires that humans know when they are interacting with AI, when content is AI-generated, and when AI is analysing their emotions or biometric characteristics.
Article 50 Compliance Checklist
1
Interaction Disclosure
Clearly inform users they are interacting with an AI system before the interaction begins, unless it is obvious from context (e.g. a clearly labelled chatbot interface).
2
Synthetic Content Labelling
AI-generated text, audio, video (including deepfakes), and images must carry machine-readable metadata and, where appropriate, a visible label identifying the content as synthetic.
3
Emotion & Biometric Disclosure
Where emotion recognition or permitted biometric categorisation is in use, inform subjects of the system's purpose and the logic involved.
4
GDPR Alignment
All disclosures must be concise, intelligible, and layered in accordance with GDPR Articles 12–14. Burying the disclosure in terms and conditions does not satisfy the obligation.
Transparency violations carry fines of up to €7.5M or 1.5% of annual global turnover, whichever is higher. This is the lightest penalty tier in the Act — but for small organisations, €7.5M can be existential. The 1.5% floor means large corporations face proportionally heavier consequences for what may seem like administrative failures.
The GPAI Model Overlay: A Separate Regulatory Track
General-Purpose AI models — systems trained on vast datasets capable of performing a wide range of tasks — operate under a parallel regulatory track rather than slotting neatly into the four-tier hierarchy. The GPAI rules became enforceable in August 2025.
Standard GPAI
All GPAI Models
✓
Maintain technical documentation
✓
Publish summaries of training data used
✓
Comply with EU copyright law
Systemic Risk GPAI
Models > 10²⁵ FLOPs
(Commission can designate below threshold)
+
Mandatory adversarial testing (red-teaming)
+
Incident reporting and risk mitigation plans
+
Mandatory energy efficiency and consumption reporting
Oversight of GPAI is centralised under the EU AI Office, supported by a Scientific Panel of independent technical experts and an Advisory Forum representing industry and civil society. The AIGP exam tests the distinction between these bodies — the AI Office supervises GPAI providers directly, while the AI Board (composed of Member State representatives) coordinates national enforcement authorities.
Enforcement Timeline and the Fine Structure
The Act's phased implementation schedule is a high-frequency exam topic. Candidates must know not just what each rule requires, but precisely when it became or becomes enforceable. Confusing the enforcement dates is a common error in scenario-based questions that specify a particular date.
Date
Milestone
Aug 2024
Act enters into force.
Feb 2025
Prohibitions (Tier 1 Unacceptable Risk) become enforceable. Nine practices banned.
Aug 2025
GPAI model rules apply. EU AI Office begins active supervision of GPAI providers.
Aug 2026
Transparency rules and most Annex III High-Risk rules apply (employment, essential services, law enforcement, migration, education, justice).
Dec 2027
High-Risk rules for Annex III Biometrics and remaining categories fully apply.
Aug 2028
High-Risk rules for Annex I product-embedded systems apply (medical devices, machinery, toys).
Prohibited Practice Violations
€35M
or 7% of global annual turnover
High-Risk Non-Compliance
€15M
or 3% of global annual turnover
Transparency / Incorrect Information
€7.5M
or 1.5% of global annual turnover
Exam Readiness: The Five Rules You Cannot Afford to Forget
The AIGP exam rewards candidates who can reason through the Act's architecture quickly and apply it to novel fact patterns. These five synthesis rules represent the most common failure points across all examination domains.
Five Rules for the Exam Room
- Risk is contextual. Classification flows from the deployment purpose, not the model architecture.
- Article 25 changes who is responsible. Trademarking or substantially modifying a system converts a Deployer into a Provider with full provider obligations.
- The Article 6(3) exception never applies to profiling. If profiling is present, high-risk classification stands regardless of how narrow the task appears.
- Documentation is the evidence of compliance. An organisation can have a perfectly fair algorithm and still be non-compliant if the Quality Management System and Technical Documentation do not exist.
- Know your governance bodies. EU AI Office supervises GPAI. AI Board coordinates Member States. Scientific Panel provides technical expertise. These are not interchangeable on exam questions.
The EU AI Act does not regulate algorithms — it regulates decisions. Every obligation in the Act traces back to a fundamental question: who is harmed, how badly, and who is accountable? Approach every exam scenario with that question first, and the risk tier answers itself.
