EU AI Act - The EU's AI Law Explained
What Europe decided while others were still debating — an overview.
Imagine an AI deciding whether you get a loan, whether your child is admitted to their preferred school, or whether you're invited to a job interview. From December 2, 2027, strict rules apply to exactly these high-risk applications in the EU.
The EU AI Act is considered the world's first comprehensive AI law – and even if you never develop AI yourself, it affects your daily life. In July 2026, the EU revised it with the "Digital Omnibus on AI" (AI Omnibus): deadlines postponed, obligations eased, two prohibitions added. This article reflects the state of November 2026: what the law regulates, when different provisions apply, and what it specifically means for you.
The Risk Level System
The EU AI Act uses a risk-based approach: the higher the potential for harm from an AI application, the stricter the requirements. Like a traffic light, the system shows four levels.
Caution
Red - Unacceptable Risk
These AI applications are completely banned. Examples: social scoring, manipulative AI, mass collection of facial images from the internet.
Caution
Orange - High Risk
Strict requirements and documentation obligations. Examples: AI that grades exams, pre-sorts job applications, or assesses creditworthiness.
Note
Yellow - Limited Risk
Transparency requirements: users must know they're interacting with AI. Example: chatbots must identify as AI.
Note
Green - Minimal Risk
No specific requirements. Examples: spam filters, AI in video games, recommendation systems.
Interactive: Which AI Falls into Which Risk Category?
You now know the four risk levels of the EU AI Act. Click on a row in the matrix to see why that use case falls into its respective category.
Risk Matrix: Use Cases Under the EU AI Act
Click a row for detail view
| Use Case | Risk Level | Requirements | Max. Penalty |
|---|---|---|---|
| Social Scoring | Prohibited | Fully banned | €35M / 7% |
| Medical AI | High Risk | Conformity assessment, CE marking, documentation | €15M / 3% |
| AI Recruitment | High Risk | Conformity assessment, CE marking, documentation | €15M / 3% |
| ChatGPT | GPAI Rules | Transparency, copyright compliance | €15M / 3% |
| Chatbot | Limited | Disclosure only | €15M / 3% |
| Spam Filter | Minimal | None | - |
What is Prohibited?
The EU AI Act bans certain AI applications outright. Eight of them have been prohibited since February 2, 2025. From December 2, 2026, two more are added: AI that generates realistic intimate images of identifiable people without their consent, and AI that generates depictions of child sexual abuse. For providers, these two bans apply if generating such content is the system's purpose, or if it can foreseeably produce such content without reliable safeguards preventing it – so an image generator with effective filters is not covered. Anyone who deliberately uses an AI system for this purpose in a professional capacity (a "deployer" in the law) also violates the ban. The AI Act does not apply to private individuals who use AI purely for themselves; however, creating or sharing such images can be a criminal offence – for depictions of child sexual abuse, throughout the EU. In the law, these are two separate points (Art. 5(1)(ba) and (bb)); the European Commission groups them as a single ninth prohibition in its overview.
Caution
The Prohibited AI Practices: 8 since 2025, 2 more from December 2026
- Manipulative AI: Systems that use subliminal, purposefully manipulative, or deceptive techniques to influence people's behavior in a way that threatens them with significant harm.
- Exploitation of vulnerabilities: AI that specifically exploits age, disabilities, or economic hardship.
- Social scoring: Rating systems that classify and disadvantage people based on their behavior or personal characteristics.
- Predictive policing: AI that predicts the risk of someone committing a crime based solely on profile data.
- Facial image scraping: Untargeted mass collection of facial images from the internet or from surveillance cameras to build facial recognition databases.
- Emotion recognition in the workplace and in schools: AI that attempts to recognize the emotions of employees or students – except for medical or safety reasons.
- Biometric categorization: AI that categorizes people by sensitive characteristics such as origin, religion, or sexual orientation.
- Real-time facial recognition: Live surveillance with facial recognition in public spaces by police (with very narrow exceptions).
- Sexual deepfakes without consent (from December 2, 2026): AI that generates or manipulates realistic nude or sexually explicit images, videos, or audio of an identifiable person without that person's explicit consent – such as so-called nudify apps.
- Depictions of child sexual abuse (from December 2, 2026): AI that generates or manipulates such material. Producing such material has long been a criminal offence in the EU; from this date, the AI systems that generate it are banned as well.
High-Risk AI Under Scrutiny
In eight sensitive areas, AI systems must meet particularly strict requirements. These are defined in Annex III of the law and apply from December 2, 2027. What matters is the specific purpose, not the area alone: an AI that pre-sorts job applications or grades exams is high-risk AI; an AI that only schedules appointments in the HR department is not. If a system performs only a narrowly limited auxiliary task and does not materially influence the decision, the provider can exempt it with a documented justification. But if it evaluates people based on their personal characteristics (profiling), it always remains high-risk AI.
High-risk AI requires, among other things: a risk management system, high-quality training data examined for biases, technical documentation, automatic logs, human oversight, and accuracy and robustness. Before market launch, the provider checks – usually itself – that everything is met, affixes the CE marking, and registers the system in the EU database.
ChatGPT & Co. - The GPAI Rules
General-Purpose AI (GPAI) – versatile AI models such as those behind ChatGPT, Claude, or Gemini – has its own rules. The obligations have applied since August 2, 2025; the European Commission has been able to impose fines since August 2, 2026. Models that were already on the market before August 2025 must comply by August 2, 2027. To help providers get there, a voluntary Code of Practice has existed since July 2025, which the Commission and the AI Board have confirmed as adequate. Signatories include OpenAI, Anthropic, Google, Microsoft, Amazon, and Mistral AI; xAI signed only the chapter on safety and security.
Note
Important: ChatGPT is NOT a High-Risk System
A common misconception: ChatGPT and similar chatbots are not automatically high-risk AI. However, they must meet transparency requirements: the chatbot must make clear that you are talking to an AI, and generated text, images, or audio must be marked as AI-generated in a machine-readable way. This has applied since August 2, 2026; systems that were already on the market before then have until December 2, 2026 for the marking. A voluntary code of practice on labelling AI-generated content has described how to do this in practice since June 2026, and guidelines from the European Commission have done so since July 2026. The code comes with three uniform EU icons for labelling deepfakes (image, audio, video) and published AI-generated text. The icons are optional; the labelling duty is not. Where the model and the chatbot come from the same provider, the Commission's AI Office has supervised the system directly since the AI Omnibus – rather than the authority in each individual EU country.
Deep dive
Standard GPAI
All GPAI models must: keep technical documentation ready for the authorities, give companies that build the model into their products the information they need, have a policy for complying with copyright law, and publish a summary of their training data. Since July 2025, the European Commission has provided a template for this summary. Exception: for open-source models without systemic risk whose weights are public, the first two duties do not apply.
Caution
GPAI with Systemic Risk
Particularly powerful models have additional obligations: model evaluations, adversarial testing (red teaming), incident reporting, and cybersecurity measures. As a rule of thumb: anyone who trains a model with more than 10²⁵ FLOPs (computing operations) must notify the Commission within two weeks – systemic risk is then presumed. The provider can rebut this presumption, and the Commission can also designate smaller models.
The Timeline - When Does What Apply?
The EU AI Act has been in force since 2024, but its rules apply in stages. In 2026, the AI Omnibus postponed some dates. Here are the key ones:
Law in Force
The AI Act officially entered into force.
Prohibitions Apply
The first eight prohibited AI practices are now banned. The AI literacy obligation has applied since then too. It was weakened in July 2026: providers and deployers must support the AI literacy of their staff, but no longer have to guarantee any specific level.
GPAI Rules Apply
The rules for GPAI models apply. EU countries must have designated their supervisory authorities.
Digital Omnibus on AI (AI Omnibus) in Force
Amending Regulation (EU) 2026/1744 postpones the high-risk deadlines, weakens the AI literacy obligation, and eases the burden on small and medium-sized enterprises (SMEs) and small mid-caps.
Transparency Obligations Apply
Chatbots must disclose that they are AI, AI-generated content must be marked in a machine-readable way, and deepfakes must be labelled. The European Commission can now also impose fines on providers of GPAI models.
Two New Prohibitions, Marking Deadline
From this date, AI systems that generate sexualised images of real people without consent, and AI-generated depictions of child sexual abuse, are also prohibited. Generative AI systems that were already on the market before August 2, 2026 must mark their output in a machine-readable way by this date.
Regulatory Sandboxes and Older GPAI Models
Each EU country must operate at least one AI regulatory sandbox in which companies can test AI under supervision. GPAI models that were already on the market before August 2, 2025 must now meet all obligations.
High-Risk Rules Apply
The rules for high-risk AI in the 8 sensitive areas (Annex III) apply. The original plan was August 2026. The EU postponed the start because the technical standards and guidelines that companies are meant to use to demonstrate compliance were not ready in time.
Embedded High-Risk AI
The rules for high-risk AI in products that already require a safety assessment today (Annex I) apply, for example to toys, lifts, or medical devices. For vehicles and machinery, their own laws are taking over the requirements step by step.
Deadline for Public Authorities
High-risk AI used by public authorities that was already on the market before the high-risk rules started must meet all requirements by this date at the latest.
Penalties for Violations
The EU AI Act provides for significant penalties – some higher than under the GDPR. For large companies, the higher of the two amounts applies; for small and medium-sized enterprises (SMEs) and start-ups, the lower one – except for the Commission's GPAI fines. Since the AI Omnibus, the lower amount also applies to small mid-caps (fewer than 750 employees and no more than EUR 150M in turnover or EUR 129M in total assets), though neither for prohibited practices nor for GPAI fines:
GDPR: A Turning Point for Data Protection with AI Impact
On May 25, 2018, the EU General Data Protection Regulation (GDPR) became applicable - a turning point for AI and data protection worldwide. As the 'mother of all data protection laws,' it replaced the outdated Directive 95/46/EC from 1995, a relic of the internet's early days. GDPR made Privacy by Design mandatory: data protection must be built into AI systems from the ground up. Its global reach was enormous - even US tech giants must comply with EU standards when processing European data. For AI, this posed a fundamental challenge: how do you explain 'black box' algorithms when GDPR demands transparency? Observers saw an incentive to develop AI with a leaner data footprint, and methods such as transfer learning gained in importance. GDPR inspired data protection laws around the world, from California to Brazil. The regulation laid the groundwork for the EU AI Act in 2024 - the step from data protection to AI regulation was a logical one.
Caution
Penalty Framework by Violation Type
| Violation | Maximum Penalty |
|---|---|
| Prohibited AI practices | EUR 35M or 7% of worldwide revenue |
| High-risk AI obligations | EUR 15M or 3% of worldwide revenue |
| GPAI model obligations | EUR 15M or 3% of worldwide revenue |
| False information to authorities | EUR 7.5M or 1% of worldwide revenue |
What Does This Mean for Me?
Note
For Citizens
More transparency in AI decisions, a right to complain to the market surveillance authority, a right to an explanation for high-risk decisions, and protection from manipulative AI and social scoring.
Note
For Businesses
Compliance requirements for high-risk AI, documentation obligations, but also legal certainty and a clear framework for innovation.
Definition
Global Impact
Like GDPR, the 'Brussels Effect' could take hold: anyone offering AI in the EU, or operating AI whose output is used here, must adapt – regardless of where the company is based.
Interactive: Should AI Be More Strictly Regulated?
You have now explored the EU AI Act in detail — its risk levels, prohibitions, obligations, and penalties. Now you can weigh the evidence yourself: click on pro and contra arguments and observe which way the scale of the regulation debate tips.
Should AI be more strictly regulated?
Click on arguments to place them on the scale. The more weight one side has, the more the scale tips.
Pro Arguments
Contra Arguments
What does this exercise show?
The AI regulation debate has no simple answer. Both sides have weighty arguments. In practice, legislators seek a middle ground between innovation and protection — e.g., the EU AI Act with its risk-based approach.
Note: The weighting of arguments is a didactic simplification. In reality, the weight depends on the specific context, use case, and societal priorities.
Key Takeaways
Note
6 Points to Remember
- The EU AI Act uses a risk-based approach: the more dangerous the AI, the stricter the rules.
- 8 AI practices have been completely prohibited since February 2025, including social scoring and manipulative AI; from December 2, 2026, sexual deepfakes without consent and AI-generated child sexual abuse material are added.
- In 8 sensitive areas (education, employment, law enforcement, etc.), particularly strict high-risk rules apply.
- ChatGPT and similar chatbots are NOT automatically high-risk - but they must be transparent.
- Prohibitions have been in effect since February 2025; the high-risk rules apply from December 2027 – in 2026, the AI Omnibus postponed their start by 16 months because the technical standards were not yet ready.
- The law applies to anyone offering AI in the EU - even companies outside Europe.