Artificial Intelligence (AI) is entering our professional and private lives. But when an algorithm causes damage, who is liable for compensation? The abandonment of the European directive on civil liability opens up a period of legal uncertainty.
Artificial intelligence is no longer a fiction: it decides, recommends and anticipates. It is being used more and more in finance, healthcare, transport and human resources. But what happens when an AI-generated decision or action causes damage? The question of civil liability is at the heart of legal debates.
An AI without legal personality
An AI, however autonomous it may be, has no legal personality. It cannot therefore be held liable. The burden rests on the natural or legal persons who design, supply or use these systems. There are three main categories:
The designer or developer, where the damage results from a design fault or insufficient testing. The supplier, if the delivered system has a defect. The user, in the event of misuse or failure to check.
“AI, however autonomous it may be, has no legal personality. It cannot therefore be held liable.” – Anne-Marie Ka
However, this allocation raises difficulties when systems evolve autonomously. Who is responsible for unpredictable behaviour?
The changing European framework
Adopted in June 2024, the European AI Regulation (AI Act) establishes a harmonised framework for the marketing and use of AI systems. It classifies uses by level of risk and imposes enhanced compliance obligations for high-risk applications, such as credit granting or financial risk management. But it does not create an autonomous liability regime.
The European Commission had proposed a directive on civil liability in AI matters in 2022, aimed at facilitating compensation actions thanks to presumptions of fault or causality. However, in 2025, this proposal was withdrawn due to a lack of political consensus. Some States considered the text premature, believing that the AI Act and the reform of the Directive on defective products were sufficient to provide a framework for the subject.
Defective products as a safety net
The new Directive 2024/493 on liability for defective products now also covers software and therefore AI systems. It introduces a presumption of defectiveness and causality, lightening the burden of proof for victims. But this approach remains imperfect in the face of evolving AI: an algorithm can adopt unforeseen behaviour after it has been put into service, weakening the effectiveness of the regime.
Towards presumed civil liability?
In the absence of a specific European regime, traditional civil law remains the basis: demonstrating fault, damage and a causal link.
Or, these elements are particularly difficult to establish in the face of autonomous AI systems.
Hence the tendency to provide for legal presumptions, such as, for example, a presumption of defectiveness in relation to defective products, as provided for in the new directive. Other presumptions could be envisaged, such as a presumption of fault, if players fail to prove compliance with the compliance obligations set out in the IA Act, or a presumption of causality. However, these mechanisms would place increased, even heavy, responsibility on AI designers and suppliers, which should be seen in the context of the considerable benefits generated by these technologies. The issue therefore remains a delicate one: we need to avoid both impunity for damage and easy legal solutions that would put the brakes on innovation.
Concrete examples of risks
Case studies are multiplying:
- An autonomous drone injuring a passer-by.
- A recruitment algorithm that discriminates against a candidate.
- A trading program that triggers a flash crash.
In each of these scenarios, AI is just a tool. But it makes identifying the person responsible and establishing the causal link more complex.
Persistent legal uncertainty
The withdrawal of the civil liability directive leaves a vacuum. Each Member State could be tempted to develop its own rules, with the risk of European legal fragmentation. In the meantime, national courts will have to adapt existing regimes on a case-by-case basis.
“By positioning itself as a trusted laboratory for AI , Luxembourg can turn these constraints into a competitive advantage, attracting international players keen on a clear and predictable framework.” – Anne-Marie Ka
Luxembourg as a laboratory of trust
For a financial and technological centre like Luxembourg, the subject is strategic. Banks, insurance companies, fintechs and AI start-ups must integrate these new requirements into their governance. Documentation, systems monitoring, traceability of algorithmic decisions: these are all tools that are becoming essential.
By positioning itself as a trusted laboratory for AI, Luxembourg can turn these constraints into a competitive advantage, attracting international players concerned about a clear and predictable framework.
Conclusion
The civil liability of artificial intelligences is still under construction. The abandonment of the European directive has shifted the burden of adaptation onto the courts and traditional systems. There is a delicate balance to be struck: protecting victims effectively while avoiding putting the brakes on innovation. It is on this compromise that, in fine, confidence in AI will depend.
