Belgium - Market Insights

The liability vacuum: Belgium’s most pressing AI challenge in 2026

As Belgium enters 2026, legal professionals and business leaders find themselves navigating an increasingly paradoxical landscape. On one hand, the EU AI Act, i.e. Regulation 2024/1689, is now firmly embedded in operational reality, with prohibitions on unacceptable-risk systems in force since February 2025 and general-purpose artificial intelligence (AI) rules effective since August 2025. On the other hand, the formal scrapping of the EU AI Liability Directive in 2025, as part of the European Commission’s Work Programme, has created a conspicuous gap that Belgium, like its European neighbours, must now confront without the anticipated harmonised safety net. Compounding this uncertainty, the European Commission’s proposed Digital Omnibus on AI (published in November 2025) introduces a “stop-the-clock” mechanism and conditional extensions that could slow down or postpone key obligations for high-risk AI systems, potentially shifting full application of high-risk requirements from August 2026 to as late as December 2027 (or August 2028 in some cases), contingent on the readiness of harmonised standards, guidelines, and support tools. While this proposal aims to provide breathing room for implementation challenges, lawyers and companies should carefully monitor its legislative progress and take these potential delays into account when planning compliance strategies, as the changes remain subject to adoption by the European Parliament and Council and could introduce temporary legal uncertainty if not finalised in time.

This is the issue keeping boardrooms and law firms awake at night. For all the regulatory architecture surrounding AI development and deployment, the fundamental question of who pays when AI causes harm remains frustratingly unresolved at the EU level. Belgium’s response to this vacuum (or lack thereof) deserves critical examination.

A federated complexity problem

Belgium’s constitutional architecture adds a distinctive layer of challenge that practitioners elsewhere may not fully appreciate. The National Convergence Plan for the Development of Artificial Intelligence, approved by the Council of Ministers on 28 October 2022 under the ambitious banner of #SmartAINation, explicitly acknowledges that national integration must proceed “in close synergy with the federated entities”. This is diplomatic language for a structural reality: AI governance in Belgium is fragmented across federal, Flemish, Walloon, and Brussels frameworks, each with distinct priorities and institutional arrangements.

Consider the practical implications:

• Export authorisations for dual-use AI technologies require coordination between the Federal Public Service Economy and regional bodies in Flanders, Wallonia, and Brussels.
• The Belgian Institute for Postal Services and Telecommunications (BIPT) has been designated as the lead supervisory authority for AI Act compliance, yet 21 separate authorities have been appointed for protecting fundamental rights.
• Regional initiatives like the Flemish AI Policy Plan (updated March 2024) and DigitalWallonia4.ai operate alongside federal strategies, creating potential inconsistencies in implementation.
• For legal advisers counselling multinational clients, this federated complexity transforms routine compliance assessments into intricate exercises in jurisdictional mapping. The designation of a competent central authority under Article 70 of the AI Act remained unclear as of mid-2025 — a telling indicator of the coordination challenges Belgium faces.

The liability gap in practice

The abandonment of the AI Liability Directive forces Belgium to rely on existing frameworks, principally the Product Liability Directive as implemented into Belgian law and general tort law principles. While the National Convergence Plan correctly observes that AI “is not used in a legal vacuum”, the existing legal architecture was not designed with autonomous, learning systems in mind.

From my perspective, this creates three acute problems for practitioners:

• First, the causation challenge. Traditional liability frameworks require demonstrable causal links between conduct and harm. When an AI system’s decision-making process involves millions of parameters trained on datasets that may themselves contain historical biases, establishing causation becomes extraordinarily difficult. Belgian courts have not yet developed consistent approaches to these evidentiary questions.
• Second, the allocation problem. Modern AI deployments typically involve multiple actors such as developers, deployers, data providers, and integrators. The EU AI Act establishes compliance obligations across this chain, but compliance with regulatory duties does not necessarily resolve civil liability. A healthcare AI system that meets all AI Act requirements might still cause patient harm; determining liability allocation among the algorithm developer, the hospital deployer, and the data provider remains legally uncertain.
• Third, the insurance gap. Belgian insurers are struggling to price AI-related risks without clear liability rules. Many professional indemnity and product liability policies contain exclusions for autonomous decision-making systems, leaving deployers potentially exposed. The regulatory sandboxes promoted under the National Convergence Plan offer controlled testing environments, but they cannot substitute for market-wide risk transfer mechanisms.

Belgium’s copyright stance: A bellwether?

Belgium’s August 2025 opposition to the EU AI Code of Practice for General-Purpose AI merits attention as a potential indicator of the country’s evolving posture. Digitalization Minister Vanessa Matz cited inadequate copyright protections, weak compensation mechanisms for creators, and unclear opt-out provisions as grounds for Belgium’s negative vote. This stance reveals a willingness to push back against EU-level compromises perceived as insufficiently protective of rights holders.

For intellectual property practitioners, Belgium’s position signals potential enforcement priorities. The interplay between AI training data and copyright — particularly for generative AI systems — remains contested territory. Belgian authorities may prove more receptive to rights holder complaints than jurisdictions adopting more permissive approaches, creating strategic considerations for AI developers operating across multiple markets.

What responsible deployment looks like now

In the absence of harmonised liability rules, prudent organisations operating in Belgium should adopt defensive postures that go beyond mere regulatory compliance:

• Comprehensive AI inventories. Maintain detailed registers of all AI systems in use, including third-party tools, with clear risk classifications aligned to the AI Act’s tiered approach.
• Enhanced contractual protections. When procuring third-party AI, negotiate explicit liability allocation clauses, indemnification provisions, and ongoing audit rights. Do not assume that provider compliance with the AI Act translates to contractual protection.
• Fundamental rights impact assessments. The AI Act mandates these for high-risk systems, but voluntary adoption for lower-risk deployments demonstrates due diligence and may prove valuable in future litigation.
• Human oversight protocols. Document meaningful human review processes, particularly for decisions affecting individuals’ rights (the Belgian Data Protection Authority’s September 2024 guidance emphasises this requirement).
• Incident response frameworks. Establish clear procedures for AI-related incidents, including reporting obligations under both the AI Act and GDPR.

The Belgian DPA’s guidance on GDPR–AI Act interplay provides a useful foundation, emphasising that data protection impact assessments must address AI-specific risks such as profiling, automated decision-making, and algorithmic bias.

The investment paradox

Belgium’s AI ambitions are not modest. Google’s announced EUR 5 billion investment in Belgian AI and cloud infrastructure, expected to create 300 jobs and expand data centres by 2027, underscores the country’s attractiveness as a European AI hub. The October 2025 designation as an AI Factory Antenna, providing access to advanced supercomputing resources, further enhances Belgium’s innovation ecosystem.

The Federal Government Agreement 2025–2029 explicitly commits to avoiding “gold plating” of EU directives (meaning Belgium intends to transpose them precisely as required, without adding stricter national rules or extra burdens) to preserve predictability, competitiveness, and a business-friendly environment, although concrete implementation details remain limited.

A critique and a prediction

Belgium’s AI governance approach reflects both the strengths and weaknesses of EU harmonisation. The country has largely avoided the temptation to create Belgium-specific rules, instead aligning closely with EU frameworks and international standards through bodies like the OECD, the Council of Europe’s Committee on Artificial Intelligence, and the Global Partnership on Artificial Intelligence. This creates predictability for international operators and positions Belgian lawyers as credible advisers on EU-wide compliance.

However, the passivity on liability questions is concerning. The reliance on existing product liability and tort frameworks assumes that Belgian courts will develop coherent AI-specific jurisprudence organically. This seems optimistic. Other jurisdictions — both within and outside Europe — are likely to establish clearer precedents first, potentially leaving Belgian businesses operating under greater uncertainty than competitors elsewhere.

My prediction: within the next few years, Belgium will face pressure to develop national guidance on AI liability allocation, likely through soft law instruments from the Data Protection Authority or sector-specific regulators rather than primary legislation. The emphasis on regulatory sandboxes in the National Convergence Plan may evolve to include liability testing frameworks, allowing controlled experimentation with different allocation models.

Conclusion: Vigilance required

For legal professionals advising on AI deployments in Belgium, the current moment demands vigilance and creativity. The regulatory architecture is largely in place; the liability framework is not. Bridging this gap requires proactive risk assessment, robust contractual structures, and close monitoring of enforcement developments across federal and regional authorities.

The opportunities are substantial. Indeed, Belgium’s positioning as an EU-compliant AI hub creates genuine competitive advantages for organisations willing to navigate the complexity. But the risks of complacency are equally real. In a field where technology continues to outpace legal frameworks, those who wait for clarity may find themselves overtaken by events.