The AI Act

Anyone in Belgium who develops, sells, or uses an AI system is now subject to the world’s first comprehensive AI code: the European AI Act (Regulation (EU) 2024/1689). The regulation entered into force on August 1, 2024, and has been fully applicable in principle since August 2, 2026, although the European legislature postponed the most stringent obligations to 2027 and 2028 in July 2026 through the so-called Digital Omnibus. Below, we explain the entire framework: what the AI Act is, who it applies to, which obligations apply and when, and what you need to do today.

What is the AI Act?

The AI Act is a European regulation. It therefore applies directly in all member states, including Belgium, without the need for transposition into national law. Its purpose is twofold: to enable the free movement of AI systems within the internal market, while ensuring that these systems are safe and respect the health, safety, and fundamental rights of individuals.

Rather than banning technology, the regulation opts for a risk-based approach. It is not the technology itself that is regulated, but its use: the greater the risk a particular application poses to people, the stricter the requirements. You may already be familiar with this model from the General Data Protection Regulation (GDPR), and the similarity is no coincidence: the AI Act also includes documentation requirements, accountability, and fines calculated based on global revenue.

What is an AI system?

The AI Act defines an AI system as a machine-based system designed to operate with varying levels of autonomy, which, once deployed, can demonstrate adaptability, and that derives from the input it receives how to generate output, such as predictions, content, recommendations, or decisions that may affect physical or virtual environments.

Three characteristics are decisive: autonomy, inferential capability, and adaptability. Traditional software that merely executes pre-programmed rules falls outside the definition. A system that independently derives from data how it arrives at a result falls under it. In practice, this means that a chatbot on an online store, a credit scoring model at a bank, and a generative language model may all fall under the regulation, each with a different set of obligations.

Who is subject to the AI Act?

The AI Act imposes obligations on the entire value chain. The two key players are the provider, who develops an AI system or places it on the market under their own name, and the deployer, who uses an AI system under their own responsibility in a professional context. In addition, rules apply to importers and distributors who place systems from outside the European Union on the market or distribute them.

The extraterritorial scope is important. The regulation applies not only to European entities, but also to providers and data controllers outside the EU when the output of their systems is used within the Union. A U.S. provider whose model is used by Belgian companies is therefore not exempt from the rules.

There are also exceptions. The AI Act does not apply to systems used exclusively for military, defense, or national security purposes; to activities involving purely scientific research and development; or to the purely personal, non-professional use by natural persons. Partial exemptions apply to AI released under a free and open-source license, unless the system is prohibited, qualifies as high-risk, or is subject to transparency obligations.

The risk-based approach: four levels

The AI Act classifies AI applications into four categories, each with its own regulatory framework.

1. Unacceptable risk: practices that are completely prohibited under the regulation.
2. High risk: systems that must meet strict technical and organizational requirements before they can be placed on the market.
3. Limited risk: systems subject to specific transparency requirements.
4. Minimal risk: the vast majority of AI applications, such as spam filters or AI in video games, are not subject to any additional requirements.

In addition, the regulation includes a separate chapter on general-purpose AI models—the models behind generative AI applications such as large language models. These fall outside the four traditional levels and are subject to their own regulatory framework.

Prohibited AI practices

Effective February 2, 2025, Section 5 of the AI Act prohibits a range of practices deemed unacceptable. These include, among others:

• AI systems that use subliminal or deliberately manipulative techniques to significantly disrupt individuals’ behavior, resulting in (potentially) significant harm;
• systems that exploit people's vulnerabilities due to their age, disability, or socioeconomic status;
• social scoring: the assessment or classification of individuals based on their social behavior or personal characteristics, resulting in discriminatory treatment;
• predicting the risk that a person will commit a crime, based solely on profiling or personality traits;
• the indiscriminate scraping of facial images from the internet or from camera footage to build facial recognition databases;
• emotion recognition in the workplace and in education, except for medical or safety reasons;
• biometric categorization that infers sensitive characteristics, such as race, political views, union membership, or sexual orientation;
• the use of real-time remote biometric identification in public spaces for law enforcement purposes, subject to strictly defined exceptions such as the search for kidnapping victims or the threat of a terrorist attack.

Since the Digital Omnibus, two additional prohibitions have been introduced, effective December 2, 2026. The first prohibits AI systems that generate or manipulate realistic intimate or sexually explicit imagery of an identifiable person without that person’s explicit consent—so-called “deepnudes.” The second targets systems that generate or manipulate child sexual abuse material. It is noteworthy that these prohibitions apply not only to the use of such systems, but also to their placing on the market and putting into service: anyone who offers the tool is subject to the prohibition themselves, without the need for an abusive user to exist first.

High-Risk AI systems

When is an AI system considered high-risk?

The regulation provides two pathways to high-risk classification. The first pathway (Article 6, paragraph 1, and Annex I) concerns AI that is incorporated as a safety component in products already covered by European product legislation, such as medical devices, machinery, elevators, toys, or vehicles. The second pathway (Article 6(2) and Annex III) lists standalone applications in which AI directly impacts people’s lives:

• biometric identification and categorization;
• critical infrastructure, such as energy and water supplies;
• education and vocational training, such as systems that determine admission or exam results;
• employment and human resources management, such as resume screening and software that supports promotions or terminations;
• access to essential private and public services, including credit scoring and risk assessment for life and health insurance;
• law enforcement;
• migration, asylum, and border management;
• the administration of justice and democratic processes.

However, this classification is not automatic. A system listed in Annex III is exempt from the high-risk regime if it does not pose a significant risk to health, safety, or fundamental rights—for example, because it performs a purely preparatory or supporting function. We discussed this exemption under Article 6(3) earlier in our article on the exception to the high-risk classification. The Digital Omnibus Regulation has further narrowed the scope: systems used exclusively for non-safety-related assistance, performance optimization, automation, convenience, or quality control are now explicitly not considered safety components. Anyone invoking an exception must document that assessment and be able to submit it upon request by the regulatory authority.

What requirements must a high-risk system meet?

For high-risk systems, the regulation imposes seven core requirements that must be met throughout the system’s entire lifecycle:

1. Risk management: a continuous process that identifies, analyzes, and mitigates risks to health, safety, and fundamental rights.
2. Data Governance: Training, validation, and test data must be relevant, sufficiently representative, and as error-free as possible, with attention paid to potential bias.
3. Technical documentation: a file demonstrating that the system complies with the regulation, prepared before it is placed on the market.
4. Logging: The system must automatically record events so that incidents can be traced afterward.
5. Transparency and operating instructions: The person responsible for operating the system must be able to understand how it works, as well as its capabilities and limitations.
6. Human oversight: The system must be designed so that a human can effectively intervene, disregard the output, or shut down the system.
7. Accuracy, robustness, and cybersecurity: the system must perform consistently and be resistant to errors and attacks. Systems that meet the Cyber Resilience Regulation (Cyber Resilience Act, or CRA) are now considered to also meet the cybersecurity requirements of the AI Act following the passage of the Digital Omnibus Act; you can read more about that regulation on our page about the CRA.

Responsibilities in the value chain

The supplier bears primary responsibility. It conducts a conformity assessment—similar to the CE marking for products—establishes a quality management system, registers the system in the European database, affixes the CE marking, and takes corrective action if the system no longer complies. For SMEs and, since the Digital Omnibus, also for small mid-cap companies, simplified documentation requirements and a more proportionate quality management system apply.

Before placing a product on the market, the importer must verify that the conformity assessment has been conducted and that the documentation is available. The distributor must verify the CE marking and the accompanying documents. Furthermore, anyone—whether a distributor, importer, or person responsible for use—who offers a system under their own name or makes a substantial modification to it becomes a supplier themselves, with all the obligations that entails.

The person responsible for the system uses it in accordance with the operating instructions, ensures that relevant input data is provided, organizes human oversight by trained personnel, retains the logs, and reports serious incidents. Employers who deploy a high-risk system in the workplace must inform the affected employees and their representatives. In addition, government agencies and certain private entities, such as banks and insurers, must conduct a fundamental rights impact assessment (FRIA) before putting the system into use. Since the passage of the Digital Omnibus Act, this assessment may cross-reference the data protection impact assessment that already exists under data protection law , thereby avoiding duplication of effort.

Transparency requirements: chatbots, deepfakes, and AI-Generated content

For AI systems with limited risk, Article 50 of the AI Act imposes transparency obligations effective August 2, 2026. The four most important ones are:

• Anyone who interacts directly with an AI system, such as a chatbot, must be informed of this, unless it is already clear from the context.
• Providers of systems that generate synthetic audio, images, video, or text must label that output as artificially generated or manipulated in a machine-readable format, for example, using a watermark.
• Users who distribute a deepfake must disclose that the content is artificially generated or manipulated. For content that is clearly artistic, creative, or satirical, an appropriate disclosure that does not interfere with the enjoyment of the work is sufficient; we discuss how this works in the case of parodies in our blog on labeling a deepfake parody.
• Anyone who publishes AI-generated text to inform the public about matters of general interest must disclose this fact, unless there is human editorial oversight and accountability.

Incidentally, it is not always enough for a chatbot to simply identify itself as a chatbot; the requirement goes beyond that single sentence, as we explained in our analysis of the transparency requirements under the AI Act. In July 2026, the European Commission also published guidelines regarding Article 50 and a code of practice on making AI-generated content recognizable, which serve as reference documents for borderline cases.

One transitional provision deserves attention: providers of generative systems that were already on the market before August 2, 2026, have until December 2, 2026, to bring the machine-readable label into compliance. Anyone releasing a new system must comply immediately.

General-Purpose AI (GPAI) models

The models underlying generative AI, such as large language models, will have their own chapter in the AI Act. The rules for these general-purpose AI (GPAI) models will take effect on August 2, 2025.

Every provider of a GPAI model must prepare and keep technical documentation up to date, make information available to providers who integrate the model into their own systems, implement a policy to comply with European copyright law, including respecting any objections that rights holders may raise against text and data mining, and publish a sufficiently detailed summary of the content used for training. You can read more about the relationship between AI training and copyright on our page about Artificial Intelligence and Copyright.

For the most powerful models, which may pose a systemic risk (the legal presumption applies to training volumes exceeding 10^25 floating-point operations), additional obligations apply: model evaluations and adversarial testing, the assessment and mitigation of systemic risks, reporting serious incidents to the AI Agency, and implementing appropriate cybersecurity measures.

To make compliance feasible, the AI agency published on July 10, 2025 a voluntary Code of Practice for GPAI Providers. The GPAI rules have been enforceable since August 2, 2026, for new models; models that were on the market before August 2, 2025, must be compliant by August 2, 2027.

AI literacy

Article 4 of the AI Act requires providers and controllers, effective February 2, 2025, to promote AI literacy among their staff. The Digital Omnibus Act softened that requirement in July 2026: companies are no longer required to ensure an adequate level of AI literacy, but must take measures to support its development. The text explicitly adds that no specific level of knowledge per person must be guaranteed. Nevertheless, a training policy and internal guidelines on the use of AI tools remain the most obvious way to implement this requirement.

Oversight and enforcement

The European level

Oversight of the AI Act is divided between the European and national levels. Within the European Commission, the European Artificial Intelligence Agency—or the AI Agency for short—is responsible for overseeing GPAI models. The Digital Omnibus has significantly expanded that role: the AI Office now also has exclusive jurisdiction over AI systems built on a GPAI model from the same provider, and over AI systems that are part of a very large online platform or a very large online search engine within the meaning of the Digital Services Act (DSA). To that end, it was given a set of tools derived directly from competition law: requests for information, on-site inspections involving the sealing of premises, binding commitments, and the authority to impose fines and penalty payments on its own.

In addition, the European Artificial Intelligence Board (the AI Board), which includes representatives from all member states, oversees the consistent application of the regulation, assisted by a stakeholder advisory forum and a scientific panel of independent experts.

The Belgian level

Each Member State must designate at least one market surveillance authority and one notifying authority. In Belgium, the Belgian Institute for Postal Services and Telecommunications (BIPT) serves as the primary market surveillance authority and central coordination point in accordance with the coalition agreement, while the FPS Economy coordinates implementation and the existing sectoral regulators retain their roles for the products listed in Annex I. However, Belgium fell behind the European timeline: as of mid-2026, the law formally establishing this regulatory framework had not yet been approved by Parliament. In practice, anyone who has a complaint about an AI system today is therefore still largely reliant on the regular courts and existing supervisory authorities, such as the Data Protection Authority when personal data is involved.

Member States must also have at least one AI regulatory sandbox operational by August 2, 2027, in which companies can develop and test their AI systems under the supervision of the regulator. This is a particularly valuable tool for SMEs and startups.

Sanctions

The AI Act's penalty framework is modeled after that of the GDPR, but goes further. There are three levels:

1. Prohibited practices: administrative fines of up to 35 million euros or 7 percent of total worldwide annual revenue, whichever is higher.
2. Violations of most other obligations, including those related to high-risk systems and transparency rules: up to 15 million euros or 3 percent of global annual revenue.
3. Providing inaccurate, incomplete, or misleading information to the authorities: up to 7.5 million euros or 1 percent of global annual revenue.

For SMEs and startups, the lower of the two amounts applies in each case, so that the fine remains proportionate to the size of the business. The AI agency may also impose penalty payments on GPAI providers and recover its oversight costs.

The Digital Omnibus: what has changed by 2026?

On July 27, 2026, Regulation (EU) 2026/1744 went into effect: the digital omnibus package on AI. The European legislature determined that the harmonized technical standards and national supervisory structures were implemented too late, resulting in a heavier regulatory burden than intended. The omnibus package therefore addresses four key areas.

First, the high-risk obligations are being phased in: they will apply as of December 2, 2027, to the stand-alone high-risk systems listed in Annex III, and as of August 2, 2028, to systems that serve as safety components in a regulated product listed in Annex I. Second, as explained above, two new prohibited practices will take effect on December 2, 2026. Third, some obligations are being eased: the AI literacy requirement has been relaxed, the definition of a safety component has been narrowed, and SMEs and small mid-cap companies will benefit from simplified documentation. Fourth, the AI Agency will be granted the enforcement powers described above.

What remains unchanged is at least as important: the prohibited practices, the GPAI rules, and the transparency requirements remain fully in effect. A postponement is therefore not an exemption. You can find a detailed discussion of the omnibus bill in our article on the deferral of high-risk obligations.

When does what apply? The timeline

• August 1, 2024: Effective date of the AI Act.
• February 2, 2025: Prohibited AI Practices and the AI Literacy Requirement.
• August 2, 2025: Rules for GPAI models, governance provisions, and penalty provisions.
• August 2, 2026: general effective date, including the transparency requirements of Article 50; GPAI rules enforceable for new models.
• December 2, 2026: the two new bans (deepnudes and child sexual abuse material); end of the transitional period for the labeling of AI-generated content by existing generative systems.
• August 2, 2027: Deadline for GPAI models that were on the market before August 2, 2025; mandatory operational AI testing environment for each Member State.
• December 2, 2027: High-risk requirements for the independent systems listed in Annex III.
• August 2, 2028: High-risk requirements for AI in regulated products listed in Annex I.
• August 2, 2030: Deadline for high-risk systems intended for use by government agencies.

The AI Act alongside other legislation

The AI Act does not stand alone. Anyone who uses AI must read the regulation in conjunction with the rest of the digital legal framework, which we explain on this site, instrument by instrument, on our overview page about the legal framework for Artificial Intelligence.

The key areas of overlap: if an AI system processes personal data, the GDPR and the AI Act apply cumulatively, each with its own framework and supervisory authority. If AI is used on an online platform, the DSA comes into play. For products with digital elements, the CRA and sector-specific product safety legislation apply. For access to and sharing of data generated or used by AI systems, the Data Act. And anyone who trains generative AI or uses its output will inevitably run into copyright issues. An integrated compliance approach, in which these frameworks are considered together, is therefore not a luxury but a necessity.

What does this mean for you?

Do you develop or sell AI systems? Start with the classification: Is your system prohibited, high-risk, subject to transparency requirements, or none of the three? Document that assessment, as the regulatory authority may request it. If your system qualifies as high-risk, use the extension until 2027 or 2028 to develop the conformity assessment, technical documentation, and quality management system; a significant modification after the new deadline will bring the system into full compliance in any case. If you offer generative AI, the relevant deadline is much closer: the labeling requirement already applies to new systems and will apply to all systems starting December 2, 2026, and on that same date, the ban on “deepnude” applications will take effect.

Do you use AI in your business? Identify which AI tools are in use within the organization, from chatbots to HR software. Verify that your chatbots and AI-generated content comply with transparency rules, establish internal usage guidelines, and provide training for employees who work with AI. If you are implementing a high-risk system—for example, for recruitment or credit assessment—plan for human oversight, logging, and, where required, a fundamental rights impact assessment, and align these with your existing GDPR documentation.

Are you an employer? The use of AI systems that evaluate, monitor, or select employees almost always falls into the high-risk category. Inform the affected employees and their representatives before the system is put into use, and take into account how it interacts with labor law and the GDPR.

Are you a government agency, or do you play a role in a regulated industry? You are subject to additional obligations, such as the registration of the use of high-risk systems and the fundamental rights impact assessment, as well as a separate deadline of August 2, 2030, for existing systems. This deadline is not tied to the omnibus act and should already be included in your multi-year planning.

Conclusion

As of August 2, 2026, the AI Act is the applicable framework for anyone in Belgium who develops, markets, or uses AI professionally. The prohibited practices, the GPAI rules, and the transparency requirements are already in full effect today; the stringent high-risk obligations will take effect in December 2027 and August 2028, with new bans on deepnude and abusive applications starting in December 2026 serving as a pivotal milestone in between. Those who classify their systems now and compile their documentation in a timely manner will not only avoid fines that can reach up to 35 million euros or 7 percent of global revenue, but will also gain the trust of customers and regulators.

---