Artificial Intelligence (AI): The Legal Framework in Belgium

Current Legal Situation as of October 26, 2026

Artificial intelligence is regulated in Belgium by the European AI Regulation (AI Act), which has been directly applicable since August 1, 2024, and sets out the rules for those who develop, sell, or use AI systems. In addition, the general rules continue to apply: the liability provisions of Book 6 of the Civil Code, the General Data Protection Regulation, copyright law, and labor law. There is no separate Belgian AI law, and the Belgian supervisory authority had not yet been legally designated as of the fall of 2026. This page provides an overview; the AI Act itself, copyright law, and labor relations each have their own page.

1. Sources of Law

The core of AI law is European. Regulation (EU) 2024/1689 of June 13, 2024, establishing harmonized rules on artificial intelligence, or simply the AI Regulation or AI Act, is directly applicable in Belgium as a regulation without the need for a transposition law. It entered into force on August 1, 2024, and will be phased in through 2030. On July 27, 2026, it was amended by Regulation (EU) 2026/1744, the digital omnibus bill on AI, which defers high-risk obligations and adds two prohibitions.

In addition to the AI Act, there are other European instruments that apply to AI systems when the circumstances warrant it: the General Data Protection Regulation (GDPR) for any processing of personal data, Directive (EU) 2024/2853 on liability for defective products, the NIS2 directive and Cyber Resilience Act for cybersecurity, the Machinery Regulation (effective January 20, 2027) for AI in machines, the Platform Work Directive for the algorithmic management of platform workers and the DSM Directive for the text and data mining that enables AI training. The relationship to the Digital Services Act and the Data Act is discussed on our pages about the Digital Services Act (DSA) and Data Act.

Under Belgian law, no law includes the term “artificial intelligence” in its title. What does apply, however, is the general legislation that governs AI applications: the Civil Code (BW) regarding liability, the Code of Economic Law (WER) on copyright, patents, and market practices, the Act of July 30, 2018 regarding the protection of natural persons with respect to the processing of personal data, the Act of April 26, 2024 that implements NIS2, the Act of August 4, 1996 regarding the well-being of employees and the Act of May 10, 2007 to combat certain forms of discrimination.

On top of that is a non-binding policy framework. In 2022, the federal government approved the National Convergence Plan for the Development of Artificial Intelligence Well, and in 2025, the Minister of Digitalization and the heads of the federal government agencies, public utility institutions, and public social security institutions signed a Charter for the Responsible Use of AI in Government Services. The Flemish government published Guidelines for the Use of Generative AI within its own departments. Such texts are not binding on citizens or businesses, but they demonstrate how the government interprets legal standards, and they are cited as a source of interpretation in the event of a dispute.

2. Scope and Definitions

According to Article 3(1) of the AI Act, an AI system is a machine-based system designed to operate with varying levels of autonomy, capable of demonstrating adaptability after deployment, and that derives from the input it receives how to generate output, such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. This definition is broad but excludes traditional software that merely executes predefined rules: the key element is the “derivation” of output from input. That same definition applies throughout Belgian law, as no Belgian law provides its own definition.

The AI Act defines the roles in the value chain. The provider is the entity that develops an AI system or has it developed and places it on the market or puts it into service under its own name (Article 3, paragraph 3). The controller is the entity that uses an AI system under its own responsibility, except for personal and non-professional purposes (Article 3(4)). A company that purchases an AI tool from a supplier and uses it for its human resources management is therefore a controller; if it substantially modifies the system or applies its own name to it, it becomes a provider itself (Article 25). In addition, the regulation governs general-purpose AI models—known as GPAI models, such as the large language models behind ChatGPT, Claude, or Gemini—with specific obligations for their providers (Article 53).

The AI Act is a regulation providing for maximum harmonization in the areas it governs. The Belgian legislature may not introduce stricter or divergent product rules for AI systems covered by the regulation; it retains jurisdiction over matters outside its scope, such as liability, criminal law, and the organization of oversight. We discussed what Belgium is and is not permitted to regulate in the blog Can Belgium still introduce its own AI rules?

3. The AI Act at a Glance

The AI Act classifies AI systems by risk. Prohibited practices (Article 5) are not permitted at all; high-risk systems (Article 6 and Annex III) must comply with requirements regarding risk management, data quality, documentation, human oversight, and cybersecurity, and undergo a conformity assessment; systems with limited risk are subject only to transparency obligations (Article 50); and all other systems are unrestricted. The prohibitions and the obligation to ensure AI literacy among staff have been in effect since February 2, 2025; the transparency rules and most other provisions have been in effect since August 2, 2026; the high-risk obligations have been postponed by the Digital Omnibus Directive to December 2, 2027, and August 2, 2028. The full overview, including the prohibitions, high-risk requirements, GPAI rules, transparency obligations, timeline, and penalties, can be found on our page about the AI Act.

4. Liability for Damages Caused by AI

There is no specific liability regime for AI, neither in Belgium nor in the European Union. The European Commission withdrew its proposed directive on AI liability—which was intended to ease the burden of proof for victims—with its 2025 Work Plan dated February 11, 2025. Damage caused by an AI system is therefore assessed under common law, which provides four avenues of recourse.

The first avenue is strict liability under Article 6.5 of the Dutch Civil Code: anyone who causes damage through their own fault must compensate for it. A company that uses an AI system negligently—for example, without the verification required by the circumstances—is committing a fault, even if the system itself is not defective. Noncompliance with the AI Act is not a basis for liability in and of itself, but a court will characterize the violation of a legal obligation as a fault, and the obligations under the AI Act thus become the standard of care against which the user is assessed.

The second approach is the strict liability of the custodian of a defective item (Article 6.16 of the Dutch Civil Code). An item is defective if it does not provide the level of safety that one may reasonably expect under the given circumstances. For a physical item with built-in AI, such as a robot or an autonomous vehicle, this is the obvious legal basis. Whether a purely software-based AI system constitutes a “thing” within the meaning of this provision is disputed in legal scholarship and has not yet been decided by Belgian case law.

The third avenue is product liability, which has been incorporated into Articles 6.41 et seq. of the Civil Code since January 1, 2025, replacing the Act of February 25, 1991. The manufacturer is strictly liable for damage caused by a defect in its product. There is an important caveat here: Article 6.42 of the Dutch Civil Code defines a “product” as any tangible movable property, excluding electricity. Standalone software—and thus also an AI model—is not covered by this definition to the letter, contrary to what is often written. This changes with Directive (EU) 2024/2853, which explicitly classifies software and AI systems as products, introduces a presumption of defect and causal link when technical complexity makes proof excessively difficult, and allows the court to order the manufacturer to produce evidence. Belgium must transpose this directive by December 9, 2026; the new rules apply to products placed on the market on or after that date. Until then, the victim of a purely software-based AI system remains dependent on liability for defects and contractual liability.

The fourth approach is contractual. Anyone who provides a service using a defective tool—and an AI system is such a tool—has committed a culpable breach of contract toward the other party, except in cases of force majeure (Article 5.230 of the Dutch Civil Code). The parties may limit their liability by contract, within the limits set forth in Articles 5.89 et seq. of the Dutch Civil Code and, with respect to consumers, within the list of unfair terms set forth in the Unfair Terms Act (WER). For those who purchase AI services, the allocation of that liability between the provider and the customer is the most important part of the contract; the standard terms and conditions of the major providers place it almost entirely on the customer. You can read about what this entails in the blog Why a standard SaaS contract is not suffucuent for AI providers and on the blog Are there contract model clauses for purchasing AI?

Furthermore, anyone whose personal data is unlawfully affected by an AI system can rely on Article 82 of the GDPR, which requires the controller and the processor to compensate for material and non-material damages.

5. AI and Intellectual Property

Copyright protects only works that are the original intellectual creation of a natural person: Article XI.170 of the WER assigns copyright to the natural person who created the work. Whether AI-generated output is protected therefore depends on the degree of human intervention; in the case of a simple prompt without further creative direction, the output generally remains unprotected. On the input side, the exceptions for text and data mining (Article XI.190, 20°, and Article XI.191/1, § 1, 7° of the WER) and the copyright policy that the AI Act imposes on providers of GPAI models (Article 53(1)(c)) determine whether a model may be trained on protected works. All of these issues—from training to output to infringement—along with the case law of the Court of Justice and the recent German rulings on LAION and Suno, are covered on our page about Artificial Intelligence and Copyright; the general framework on our page about the intellectual property rights.

For patents, the rule is simpler. The inventor must be a natural person: Article XI.16 of the WER requires that the inventor be identified in the application, and Article XI.9 of the WER presumes that the applicant is the inventor. The European Patent Office confirmed in the DABUS Decisions J 8/20 and J 9/20 dated December 21, 2021, stating that a machine cannot be an inventor within the meaning of the European Patent Convention. However, an invention created with the aid of AI remains patentable in the name of the person who made it; see the blog Can an AI system be the inventor of a patent?

6. AI and Data Protection

As soon as an AI system processes personal data—whether during training, use, or in its output—the GDPR applies alongside the AI Act. The Dutch Data Protection Authority emphasizes that both frameworks complement each other and must be applied together. The traditional principles carry greater weight in the context of AI than elsewhere: the legal basis for each processing stage separately, whereby the scraping of training data constitutes a separate processing operation that requires its own legal basis; the principle of data minimization as opposed to models that are fed as much data as possible; transparency versus the opacity of the model; and the duty to provide information under Article 14 of the GDPR when the data has not been obtained directly from the data subject.

Article 22 of the GDPR remains the provision most frequently invoked today in relation to automated decisions: anyone subject to a decision based solely on automated processing that produces legal effects or similarly significant effects on them has the right to human intervention, to express their point of view, and to challenge the decision. Unlike the high-risk obligations under the AI Act, this provision is already fully in effect.

A data protection impact assessment (Article 35 of the GDPR) is required when processing involving new technology poses a high risk, which is almost always the case with AI systems that make decisions about individuals. For high-risk systems, this is supplemented by the fundamental rights impact assessment under Article 27 of the AI Act for government entities and certain private data controllers. The output of a model may also contain personal data, and a fabricated fact about an existing person constitutes inaccurate personal data, which the data subject may request to be rectified. The complete framework can be found on our page about privacy and data protection.

7. AI in the Workplace

The use of AI in the employment relationship affects three regulatory frameworks at once. The AI Act classifies AI systems used for the recruitment, selection, evaluation, promotion, and dismissal of employees, as well as for the assignment of tasks based on behavior or personal characteristics, as high-risk (Annex III, 4), prohibits emotion recognition in the workplace (Article 5, 1)(f)), and requires employers to inform employees and their representatives before implementing a high-risk system (Article 26, 7). Labor law further incorporates the Welfare Act of August 4, 1996, which requires employers to conduct risk assessments for new technologies as well, and collective bargaining agreements No. 39 on the introduction of new technologies, No. 68 on video surveillance, and No. 81 on the monitoring of electronic online communication data, which require the employer to consult with and inform employees before implementing AI-driven monitoring. The anti-discrimination laws of May 10, 2007, apply in full to algorithmic decisions: a system that, due to its training data, disadvantages women or older candidates constitutes indirect discrimination for which the employer is liable, even without intent. The Platform Work Directive, which Belgium must transpose by December 2, 2026, adds the first European rules on algorithmic management. Details for each phase of the employment relationship can be found on our page about Artificial Intelligence and the Employment Relationship; the obligation to make staff AI-literate, in the blog How can I ensure that my organization becomes AI-literate?

8. Cybersecurity of AI Systems

High-risk systems must maintain an appropriate level of cybersecurity throughout their entire lifecycle and be resistant to attempts by third parties to manipulate their operation (Article 15 of the AI Act). The regulation specifically names AI-specific attacks: data poisoning, in which training data is manipulated, and adversarial examples, in which the input is distorted in such a way that the model makes incorrect decisions. This obligation is in addition to general cybersecurity legislation. Companies and government agencies that fall under NIS2 and are subject to the Belgian law of April 26, 2024, must incorporate their AI systems into their risk management and incident reporting; financial institutions do the same under DORA. An AI system that is marketed as a product with digital elements falls under the Cyber Resilience Act, and the AI Act stipulates that compliance with the essential requirements of that regulation constitutes compliance with its own cybersecurity requirement.

9. Oversight and Enforcement in Belgium

The AI Act requires each Member State to designate at least one notifying authority and one market surveillance authority (Article 70). As of the fall of 2026, Belgium has not yet codified this requirement into law. The FPS Economy, which coordinates implementation, is working on a preliminary draft that would designate the Belgian Institute for Postal Services and Telecommunications (BIPT) as the central market surveillance authority, alongside sectoral regulators; there is still no agreement between the federal government and the regions regarding the oversight of regional competences such as education and employment services, as reported by VRT News on August 2, 2026. In the meantime, the Data Protection Authority exercises indirect oversight whenever personal data is involved, and sectoral regulators retain jurisdiction in their respective areas. For GPAI models, the European Commission’s AI Office itself serves as the supervisory authority.

The absence of a Belgian supervisory authority does not suspend these obligations. The regulation applies; civil courts enforce it in liability and contractual disputes, and the labor court may classify a violation of it as an unfair market practice. Administrative fines may only be imposed once Belgian law has designated the competent authority, but their maximum amounts are set forth in the regulation (Article 99): up to 35 million euros or 7 % of global annual turnover for prohibited practices, up to 15 million euros or 3 % for the other obligations, and up to 7.5 million euros or 1 % for providing false information to the authorities, with the higher amount applying in each case for companies.

10. AI in the Judicial System and the Legal Profession

AI systems that assist a court in examining and interpreting facts and the law are high-risk (Annex III, 8). The decision itself remains with humans. The Court of Cassation devoted its opening address for the judicial year on September 1, 2026, entirely to AI and the justice system; our analysis of it can be found on the blog Will artificial intelligence eventually replace judges and lawyers?

For attorneys and parties, the responsibility for what they submit remains unchanged. Belgian judges have already sanctioned those who included fabricated case law generated by a language model in their briefs, with penalties ranging from a reprimand to a fine for vexatious and reckless litigation; see the blog What are the risks if AI-generated legal rulings end up in your court briefs? The Flemish Bar Association issued guidelines on the use of AI by attorneys, which focus on attorney-client privilege, the duty of oversight, and client information. Anyone who uses AI tools in a law firm or legal service must review the supplier’s data processing terms in advance: what data is used to train the model, where it is stored, and who has access to it.

11. Current Developments

Three key issues will shape the coming year. The first is the Belgian implementing law for the AI Act, which designates the BIPT and divides supervisory responsibilities between the federal government and the regions; until that law is enacted, enforcement remains limited to the civil courts and existing regulatory authorities. The second is the transposition of Directive (EU) 2024/2853 by December 9, 2026, which brings software and AI systems under product liability and eases the burden of proof for victims. The third is further simplification through the Digital Omnibus, the AI section of which has already been codified in Regulation (EU) 2026/1744, while the sections on data protection and cybersecurity are still under negotiation. In addition, European codes of practice for GPAI models and for AI content are currently being developed, and by August 2, 2027, every Member State will have an AI testing environment for companies that wish to develop their systems under the supervision of the regulator.

Read more on ICT Legal Guide

How our lawyers can assist you

ICT Rechtswijzer is the knowledge platform of the Intellectual Property, IT Law, AI Law, Data Protection, and Media Law department at Everest Advocaten, led by Joris Deene, Esq. Our attorneys assist businesses and government agencies in classifying their AI systems under the AI Act, drafting contracts with AI suppliers, and resolving disputes arising from the use of AI.

Specifically, we determine which risk category your system falls into and what obligations arise from that as of which date; we negotiate and revise the license and SaaS agreements with your vendor; we draft your internal AI policy and your transparency statements; we conduct the data protection impact assessment and the fundamental rights impact assessment, and we represent you in disputes regarding liability, copyright, or discrimination caused by an AI system.

Contact

Questions? Need advice?
Contact Attorney Joris Deene.

Phone: 09/280.20.68
E-mail: joris.deene@everest-law.be

Topics