# EU AI Act guide: what is banned, deadlines and fines

> A guide to Regulation (EU) 2024/1689 (Artificial Intelligence Act) as amended by Regulation (EU) 2026/1744. Verified as of 22 September 2026 on EUR-Lex and in the Czech eKLEP system. Prepared by CIAD, the Czech Institute for AI and Data. Web version: https://ciad.cz/en/ai-act/ · Consolidated text: https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:02024R1689-20260727

This text is not legal advice. Only the text published in the Official Journal of the EU is binding.

## What is the AI Act?

The AI Act is Regulation (EU) 2024/1689 of the European Parliament and of the Council, which sets rules for developing and using artificial intelligence across the EU. It applies directly, without national legislation, and covers companies that develop AI as well as those that only use it. The heaviest penalties are for prohibited practices, such as recognising employees’ emotions or social scoring: fines of up to EUR 35 million or 7 % of worldwide annual turnover.

- **What it is:** Regulation (EU) 2024/1689, the Artificial Intelligence Act. Directly applicable across the EU, in force since 1 August 2024, amended in July 2026 by Regulation (EU) 2026/1744.
- **Who it covers:** Companies and public bodies that develop AI and those that only use it, including companies outside the EU when the output is used in the EU. It does not cover purely personal use, systems used exclusively for military and defence purposes, or systems developed solely for scientific research and development (Art. 2).
- **Who enforces it in Czechia:** The Czech act designating the supervisory authorities has not been adopted yet; the government has not discussed the draft submitted on 26 June 2026. Under the draft, supervision will be shared by the Czech Telecommunication Office (ČTÚ), the Czech National Bank (ČNB) and the Office for Personal Data Protection (ÚOOÚ). The obligations of the regulation apply directly regardless.

## Does it apply to your company?

Obligations depend on the purpose of use. A general assistant such as ChatGPT is not high-risk in itself; its use becomes high-risk when, for example, it screens job applicants.

| Sector | What applies to you | High risk |
|---|---|---|
| E-commerce and customer service | Disclosure for a chatbot you developed or had developed under your name (Art. 50(1)), labelling of deepfakes in advertising (Art. 50(4)), the ban on harmful manipulation (Art. 5(1)(a) and (b)), AI literacy. | Usually not high-risk. |
| HR and recruitment agencies | Screening CVs and evaluating candidates and performance falls under Annex III point 4. Recognising the emotions of candidates and employees is already prohibited. | High-risk from 2 Dec 2027 (Art. 26). |
| Banks, insurers, lending | Creditworthiness assessment of natural persons and pricing of life and health insurance fall under Annex III point 5. In addition, a fundamental rights impact assessment (Art. 27). Supervision by the Czech National Bank under the draft act. | High-risk from 2 Dec 2027. |
| Marketing and agencies | Labelling of deepfakes (Art. 50(4)). AI-generated text must be labelled only on matters of public interest without editorial control. From 2 Dec 2026 intimate deepfakes are prohibited (Art. 5(1)(ba)). | Usually not high-risk. |
| AI developers and SaaS | Provider role: chatbot disclosure and machine-readable marking of outputs (Art. 50(1) and (2)), safeguards against prohibited outputs, and for high-risk uses conformity assessment, documentation, CE marking and registration. Relief for small and medium-sized enterprises. | Depends on the purpose of the product. |
| Manufacturers of regulated products | AI as a safety component of a product under Annex I, for example in medical devices, lifts or toys. | High-risk from 2 Aug 2028. |
| Schools and education | Admission, assessment of learning outcomes and proctoring of exams fall under Annex III point 3. Recognising students’ emotions is prohibited. | High-risk from 2 Dec 2027. |
| Municipalities and public authorities | Eligibility for benefits and public services falls under Annex III point 5. Fundamental rights impact assessment (Art. 27), registration. Legacy systems by 2 Aug 2030 at the latest. Under the Czech draft act, fines are capped at CZK 10 million. | High-risk from 2 Dec 2027. |

## When does it apply?

- **2 Feb 2025 · Prohibitions and AI literacy** (Already applies, Art. 113(a) and (b)): Prohibited practices under Art. 5 and the obligation to take AI literacy measures under Art. 4. From 2 Aug 2025 also the rules for general-purpose AI models and the penalty framework.
- **2 Aug 2026 · Transparency** (Already applies, Art. 113, Art. 111(4)): Disclosure for chatbots, labelling of deepfakes and informing people about emotion recognition under Art. 50. From 2 Dec 2026 two prohibitions are added (intimate deepfakes and child sexual abuse material) and older content generators must mark their outputs in a machine-readable way.
- **2 Dec 2027 · High-risk systems** (Postponed by the amendment, Art. 113(c)): Obligations for high-risk uses under Annex III, for example recruitment or creditworthiness assessment. For AI in products under Annex I only from 2 Aug 2028.

## What does your company have to do?

- **Every company that uses AI:** Do not use AI for prohibited practices (Art. 5), take AI literacy measures (Art. 4), and from 2 Aug 2026 label published deepfakes and inform people about emotion recognition (Art. 50(3) and (4)).
- **A company with a high-risk use:** From 2 Dec 2027 the deployer obligations under Art. 26: use in line with the instructions, human oversight, control of input data, logs kept for at least six months, informing employees and affected people.
- **A company that develops AI or sells it under its own name:** Chatbot disclosure and machine-readable marking of generator outputs (Art. 50(1) and (2)). For high-risk systems the requirements of Art. 9 to 15, conformity assessment, CE marking and registration. A company also becomes a provider when it substantially modifies a high-risk system or places it on the market under its own name (Art. 25).

### What you must never do (Art. 5)

- Subliminal, manipulative or deceptive techniques that materially distort a person’s decision-making and cause, or are reasonably likely to cause, significant harm to them or others.
- Exploiting vulnerabilities due to age, disability or a social or economic situation, causing actual or likely significant harm.
- Social scoring that leads to detrimental treatment in an unrelated context or treatment that is disproportionate. Applies to companies and public authorities alike.
- Assessing the risk that a person will commit a criminal offence based solely on profiling or personality traits.
- Facial recognition databases built from untargeted scraping of images from the internet or CCTV.
- Inferring emotions in the workplace and in education institutions, except for medical or safety reasons.
- Biometric categorisation by race, political opinions, trade union membership, religious or philosophical beliefs, sex life or sexual orientation.
- Real-time remote biometric identification in publicly accessible spaces for law enforcement, outside narrow exceptions.
- From 2 Dec 2026: realistic intimate images of an identifiable person without their consent, and child sexual abuse material.

### What to have ready before an inspection

- An inventory of AI tools: who uses them, for what, with which data and in which role the company acts.
- A check that no tool serves a prohibited practice.
- AI literacy measures by role and a record of them. No certificate is required.
- Disclosure for your own chatbots and labelling of deepfakes.
- Classification of high-risk uses and a plan to comply by 2 Dec 2027.
- Supplier contracts that secure the instructions for use and the information you need.
- A named person responsible for keeping the records.

## What are the fines and for what?

| Infringement | Maximum fine | For example | Provision |
|---|---|---|---|
| Prohibited practices | EUR 35 million or 7 % | Software evaluates the emotions of employees or candidates; a company refuses customers based on a score derived from their social media behaviour. | Art. 99(3) |
| Obligations of operators and transparency | EUR 15 million or 3 % | Your own chatbot does not disclose that it is AI; an unlabelled deepfake in advertising; from 2 Dec 2027 screening candidates without human oversight or selling a high-risk system without conformity assessment and CE marking. | Art. 99(4) |
| Incorrect information to authorities | EUR 7.5 million or 1 % | Incorrect, incomplete or misleading answers to a request from an authority. | Art. 99(5) |
| Providers of general-purpose AI models | EUR 15 million or 3 % | Applies only to model makers. The fine is imposed by the European Commission. | Art. 101 |

For companies the higher of the two amounts applies, for small and medium-sized enterprises the lower (Art. 99(6)). The percentage is calculated from worldwide annual turnover for the preceding financial year. The regulation sets no separate fine for missing AI literacy.

### Beyond fines and during an inspection

- **Corrective action, then a ban** (Art. 79, Art. 83): If an authority finds non-compliance in a system presenting a risk, it requires corrective action within at most fifteen working days. If none follows, it prohibits or restricts placing the system on the market or putting it into service, or ensures its withdrawal. The request itself does not rule out a fine.
- **What an authority may request** (Art. 74(12) and (13)): Documentation and, for high-risk systems, training and testing data. Source code only on a reasoned request, when it is necessary to assess conformity and other means have been exhausted.
- **What reduces a fine** (Art. 99(7)): The authority takes into account, among other things, cooperation, remediation, the measures in place and whether you reported the infringement yourself. It also considers fines for the same conduct under other laws, for example the GDPR.
- **Complaints** (Art. 85): Anyone with grounds to believe an infringement has occurred may lodge a complaint with a market surveillance authority: a customer, an employee or another company.

## Frequently asked questions

### Does the AI Act apply to a company that only uses ChatGPT or Copilot?

Yes, but to a limited extent. The company must take AI literacy measures under Article 4, must not use AI for prohibited practices and must label published deepfakes. High risk depends on the purpose: if the tool is used, for example, to screen job applicants, that use falls under Annex III point 4 and under Art. 25(1)(c) the company becomes the provider of a high-risk system.

### Does the AI Act apply to small companies?

Yes. The basic obligations do not depend on size. For small and medium-sized enterprises, however, the fine cap is the lower of the two amounts (Art. 99(6)), and smaller providers get relief, for example simplified technical documentation.

### Is AI literacy training mandatory?

What is mandatory is taking measures to support AI literacy under Article 4. The regulation prescribes no format, length or certificate, and after the amendment it expressly does not require guaranteeing a particular level for each individual. Training with a record of attendance is the most common and best documented measure.

### Does all AI-generated content have to be labelled?

No. Providers of generators must mark outputs in a machine-readable way. A company that publishes content must label deepfakes and text on matters of public interest unless it has undergone human editorial review for which someone holds editorial responsibility. Ordinary text reviewed by a person, for which the company is responsible, does not need an AI label under the AI Act.

### Has anyone been fined under the AI Act yet?

As of 22 September 2026 we found no publicly known case of a fine imposed under the AI Act in the EU. The European Commission states that the enforcement powers of the AI Office and national authorities apply from 2 August 2026.

## Detailed answers

### When do AI Act obligations apply? Deadlines after the 2026 amendment

https://ciad.cz/en/answers/ai-act-compliance-deadlines-february-2025-to-august-2027/

The AI Act, Regulation (EU) 2024/1689, entered into force on 1 August 2024 and applies in stages. The Article 5 prohibitions and the Article 4 AI literacy obligation have applied since 2 February 2025. The rules for general-purpose AI models and the penalty framework have applied since 2 August 2025. Most of the rest of the regulation, including the Article 50 transparency rules, applies from 2 August 2026. Two new prohibitions, on non-consensual intimate deepfakes and child sexual abuse material, apply from 2 December 2026. Regulation (EU) 2026/1744 postponed the high-risk obligations to 2 December 2027 for Annex III uses and to 2 August 2028 for products under Annex I.

### AI training for companies: content, roles, and the Article 4 obligation

https://ciad.cz/en/answers/ai-training-for-companies-content-roles-and-the-article-4-obligation/

Corporate AI training is ordered for two different reasons, and conflating them produces training that serves neither well. The first is productivity: people should be able to use available tools competently and know their limits. The second is a legal one: Article 4 of the AI Act, as replaced by Regulation (EU) 2026/1744, requires providers and deployers of AI systems to take measures supporting the AI literacy of staff and other people who operate or use the systems on their behalf, proportionate to their role, knowledge, and the context of use. It does not require a specific format, does not require a guaranteed literacy level for each person, and does not require a certificate. A well-designed program covers different depth for different roles and produces a record showing a real measure took place, which is CIAD practice for demonstrating diligence, not itself the letter of the legal requirement.

### AI training certificates: what the certificate actually means

https://ciad.cz/en/answers/ai-training-certificates-what-the-certificate-actually-means/

A certificate means a specific provider confirms a specific person attended a specific course. Nothing more follows from it automatically. As of August 2026, there is no state or EU body that accredits AI-literacy certificates, so the weight of any given certificate rests on the reputation of the provider that issued it, not on a regulatory scheme. The AI Act's Article 4, as replaced by Regulation (EU) 2026/1744, requires organisations to take measures supporting AI literacy; it does not require a certificate, does not define a certification scheme, and the European Commission's own guidance confirms no certificate is needed. A certificate can be useful internal evidence that a measure took place. It is not, by itself, legal proof of compliance.

### How to choose an AI training provider for your company

https://ciad.cz/en/answers/how-to-choose-an-ai-training-provider-for-your-company/

Price is not the first criterion, and the biggest differences between providers rarely show up in a price list. Check five things instead: whether the provider distinguishes content by role rather than teaching everyone the same session; whether the training builds on your actual tools and processes or is a generic presentation delivered unchanged to every client; whether the content and legal framing are current, given how recently Article 4 of the AI Act was amended; whether the provider gives you real evidence of what was covered and to whom, not just a certificate; and who is actually delivering the session, since a provider's marketing material and its actual instructor are not always the same standard.

### What an internal AI policy must cover

https://ciad.cz/en/answers/what-an-internal-ai-policy-must-cover/

Internal AI policy must define a list of approved tools, rules for data classification before input into models, explicit prohibitions (e.g., entering personal data into public chatbots), a process for human verification of outputs, clear assignment of liability for damages, and deadlines for regular document review.

### What an AI system audit actually tests

https://ciad.cz/en/answers/what-an-ai-system-audit-actually-tests/

An AI system audit does not test whether a model is 'good'. It tests whether the organisation deploying or providing it can show a working risk-management process, adequate data governance, technical documentation, logging that would let an incident be reconstructed, and a real human-oversight mechanism, proportionate to the system's risk. For a high-risk AI system under the EU AI Act these are legal obligations, not best-practice suggestions. For lower-risk systems an audit still checks the same categories, scaled to what is actually at stake if the system fails or is misused.

## Sources

- EUR-Lex: [Regulation (EU) 2024/1689, consolidated text of 27 July 2026](https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:02024R1689-20260727)
- EUR-Lex: [Regulation (EU) 2026/1744 (Digital Omnibus on AI)](https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202601744)
- European Commission: [Guidelines on prohibited AI practices](https://digital-strategy.ec.europa.eu/en/library/commission-publishes-guidelines-prohibited-artificial-intelligence-ai-practices-defined-ai-act)
- European Commission: [Guidelines on the definition of an AI system](https://digital-strategy.ec.europa.eu/en/library/commission-publishes-guidelines-ai-system-definition-facilitate-first-ai-acts-rules-application)
- European Commission: [AI literacy: questions and answers](https://digital-strategy.ec.europa.eu/en/faqs/ai-literacy-questions-answers)
- European Commission: [Enforcement of the AI Act](https://digital-strategy.ec.europa.eu/en/policies/enforcement-ai-act)
- European Commission: [Market surveillance authorities under the AI Act](https://digital-strategy.ec.europa.eu/en/policies/market-surveillance-authorities-under-ai-act)
- Government of the Czech Republic: [Draft Czech act on artificial intelligence in the eKLEP system (in Czech)](https://odok.gov.cz/portal/veklep/material/KORNDLSJSEUC/)