EU AI Act

Article 3 is the dictionary of the Act. Two definitions matter most.

First, 'AI system' is defined broadly: a machine-based system that, with varying degrees of autonomy, may exhibit adaptiveness after deployment, and that infers from inputs how to generate outputs (predictions, content, recommendations, or decisions) that can influence physical or virtual environments. The Commission has explicitly clarified that simple deterministic software (e.g. spreadsheet formulas, rule-based systems without inference) is NOT in scope.

Second, the 'provider' vs 'deployer' distinction runs through the whole Act. A provider develops the AI system or has it developed and places it on the market under its own name. A deployer uses an AI system in a professional context. Different obligations attach to each — and a single company can be both at once (e.g. a fintech that builds and uses its own credit-scoring model).

Article 5 lists AI practices the EU has decided are too dangerous to ever permit. They are banned outright — no risk assessment will save you. There are 8 prohibited categories.

The most consequential bans for businesses are: (1) AI that uses subliminal or manipulative techniques to materially distort behaviour and cause harm; (2) AI that exploits vulnerabilities of specific groups (children, people with disabilities, those in poverty); (3) social scoring by public authorities; (4) untargeted scraping of facial images from the internet or CCTV to build face recognition databases; (5) emotion recognition in workplaces and schools; (6) biometric categorisation to infer race, political views, religion, or sexual orientation; and (7) most real-time remote biometric identification in public spaces by law enforcement (with narrow exceptions).

Fines for Article 5 violations are the highest in the Act: up to €35M or 7% of worldwide annual turnover, whichever is higher.

Article 6 is the gate to the heaviest part of the Act. An AI system is 'high-risk' if it either: (a) is a safety component of a product covered by the EU harmonisation legislation listed in Annex I (e.g. medical devices, machinery, toys, in-vitro diagnostics) AND requires a third-party conformity assessment under that legislation; OR (b) falls into one of the use-case categories in Annex III.

Annex III lists 8 high-risk areas: biometrics, critical infrastructure, education and vocational training, employment and workforce management, access to essential private and public services (including credit-scoring and welfare benefits), law enforcement, migration and border control, and administration of justice and democratic processes.

There is a 'lite' carve-out (Article 6(3)): if an Annex III system 'does not pose a significant risk of harm to the health, safety or fundamental rights of natural persons' — for example, it performs a narrow procedural task or improves the result of a previous human activity — the provider may document that conclusion and avoid full high-risk obligations. The Commission was meant to issue guidelines on this carve-out by Feb 2026.

If your AI system is high-risk, you must establish, implement, document, and maintain a risk management system. It is not a one-off compliance exercise — it runs across the entire lifecycle of the system and must be updated regularly.

The minimum process: identify and analyse known and reasonably foreseeable risks; estimate and evaluate risks that may emerge when the system is used as intended or under conditions of reasonably foreseeable misuse; evaluate other risks discovered through post-market monitoring; adopt risk-management measures that eliminate or mitigate risks to an acceptable level.

Special attention is required for impact on persons under 18 and other vulnerable groups. The system as a whole — when considered with the risk-management measures — must achieve 'acceptable risk' (Article 9(3)).

Training, validation, and testing datasets used for high-risk AI must meet quality criteria. They must be relevant, sufficiently representative, free of errors to the extent feasible, and complete for the intended purpose. Datasets must take into account the geographical, contextual, behavioural, or functional setting in which the system will be used.

For systems that use techniques to detect or correct bias, providers may process special-category personal data (Article 9 GDPR) under strict conditions — including a documented necessity, deletion after bias detection, and no transfer to third parties (Article 10(5)).

Data-governance practices must cover design choices, data collection and preparation, annotation, labelling, examination for biases that may affect health, safety, or fundamental rights, and gap-filling.

Article 13 says: tell your customers (deployers) enough about the system so they can use it safely. You must provide instructions for use that include the system's identity and contact details, intended purpose and expected accuracy, known limitations, training data characteristics in summary, technical capabilities and features needed to interpret outputs, human oversight measures, expected lifespan, and maintenance / updates.

This is the article that creates the documentation that ends up in your customer's compliance file. It is also what enables Article 26 deployer obligations to be met — without this information, the deployer can't comply.

High-risk AI systems must be designed so a human can effectively oversee them. The Act lists what 'effective oversight' means: the human can understand the system's capacities and limitations, monitor its operation, detect anomalies and dysfunctions, correctly interpret outputs, decide not to use the output, and intervene or stop the system.

For remote biometric identification (Article 14(5)), the deployer cannot act on an identification unless two natural persons have separately verified and confirmed it — a 'two human signoff' rule.

Deployers — not just developers — have direct obligations. You must use the system in accordance with the provider's instructions, ensure input data is relevant and representative, monitor the system and inform the provider of risks, suspend use when you suspect the system poses risk, keep logs (for at least 6 months), inform affected workers when you use a workplace AI system, and complete a fundamental rights impact assessment for some deployments (Article 27).

This matters because most companies are deployers, not providers. If you license an AI tool and use it for hiring, that tool's vendor handled the provider obligations — but Article 26 puts duties directly on you.

Article 50 covers the 'limited risk' tier — AI systems that aren't high-risk but interact with people directly enough that disclosure is warranted.

Four categories: (1) chatbots — users must be informed they are interacting with an AI unless that's obvious; (2) emotion recognition or biometric categorisation systems — the natural persons exposed must be informed; (3) deepfakes — content depicting real people, places, or events that has been artificially generated or manipulated must be clearly labelled; (4) generative text used to inform the public on matters of public interest — must be disclosed as AI-generated unless human-reviewed and a publisher takes editorial responsibility.

Labels on AI-generated images, audio, and video should be machine-readable where technically feasible (think: C2PA-style watermarking).

Chapter V deals with general-purpose AI models — what most people call 'foundation models'. Every GPAI provider has baseline obligations: maintain technical documentation, provide information to downstream providers, comply with EU copyright law, and publish a sufficiently detailed summary of the training data.

If the model exceeds the systemic-risk threshold — currently 10^25 floating-point operations of training compute — the obligations escalate sharply: model evaluations (including adversarial testing / red-teaming), risk assessment and mitigation, serious-incident reporting to the AI Office, cybersecurity protection for the model and physical infrastructure, and notification to the Commission within 2 weeks of crossing the threshold.

Providers of open-source GPAI models get a partial exemption from the documentation and downstream-information obligations — but NOT from the systemic-risk obligations if they cross the compute threshold.

Three penalty tiers. Tier 1 (the highest): violation of the Article 5 prohibitions — up to €35,000,000 or, if the offender is a company, 7% of worldwide annual turnover, whichever is higher. Tier 2: violation of other AI Act obligations (Articles 16-29 provider/deployer duties, GPAI duties, etc.) — up to €15,000,000 or 3% of worldwide annual turnover. Tier 3: supplying incorrect, incomplete, or misleading information to authorities — up to €7,500,000 or 1% of turnover.

SMEs (including startups) get a more favourable calculation: the lower of the absolute amount and the percentage applies, not the higher. National authorities also have discretion to lower fines proportionate to economic viability.