The proposal for an AI liability directive (AKA the “AI Liability Directive”) has been retracted. This appears from the 2025 Commission Work Programme.[1] The reason for retracting the proposal, which was introduced by the European Commission on 29 September 2022, is that there is ‘no foreseeable’ agreement between the European co-legislators.[2] The European Commission stated that it is examining whether another legislative proposal should take the place of the AI Liability Directive, or whether a different approach should be followed.
What was regulated in the AI Liability Directive?
The purpose of the legislative proposal was – for the improvement of the EU’s internal market – to lay down uniform rules for non-contractual civil liability for damage caused as a result of the involvement of AI systems. Since 1 August 2024, the AI Regulation has been in force, which regulates the marketing and the use of – mainly ‘high-risk’ – AI systems. This Act includes safeguards to limit the risks to security and fundamental rights related to the deployment of AI systems. However, these safeguards do not suffice to eliminate these risks fully. Moreover, the AI Regulation lacks concrete provisions for the recovery of loss resulting from defective AI systems.
That is the reason why the AI Liability Directive focused on liability for all forms of loss (incurred by natural persons and legal persons) caused by AI systems. Examples are situations in which an individual incurs a financial loss or damage to their reputation as a result of incorrect output by an AI system, for example because the security measures were not implemented correctly, or because there has been discrimination.
The legislative proposal was in line with similar laws already existing at EU level – such as the EU Product Liability Directive – that offer protection against damage incurred due to other technologies. These laws depart from the principle of strict liability, which means that it is immaterial to the recoverability of the loss whether the loss results from negligence. This Product Liability Directive – incorporated in the Civil Code in the Netherlands – may also be of assistance in relation to AI-driven products. However, its effect is limited to the producers of the defective products (barring some exceptions). The AI Liability Directive extended this liability to users of AI systems. In addition, it introduced two tools to make it easier for injured parties to prove causality between the loss incurred and the act causing the loss (by the AI system), namely (i) the ability to obligate the party responsible for the AI system to disclose the evidence, and (ii) the assumption that a causal link is presumed to exist.
i. Disclosure of evidence by deployer of AI system
The party that has developed the AI system (the provider) or the party that uses the AI system (the deployer) could be ordered to disclose evidence. This situation would only arise if the AI system suspected of having caused the loss qualified as ‘high-risk’ under the AI Regulation and if the claimant had already asked for, but not received, the information earlier. Moreover, the claimant had to ‘make it plausible’ that a loss was incurred through the AI system concerned. This way, general ‘fishing expeditions’ are prevented. The party sued could rebut the presumption of the loss caused by the designated AI system.
ii. Presumption of causality
Given the complexity of AI systems (and their functioning), the proposal provided for a – rebuttable – presumption of causality between the fault on the one hand (and the output generated by the AI system), and the inability of the AI system to generate this output that caused the damage concerned on the other hand. Besides, the claimant would still have to prove that the damage was caused by a breach of an obligation under EU law or national law (including a failure to comply with the AI Regulation), and that, given the circumstances of the case, the conduct[3] of the party sued had likely influenced the occurrence of the damage and that the (absent) output of the AI system had led to the damage.
Consequences for Practice (in the Netherlands)?
Does the foundering of the proposal for the AI Liability Directive mean that persons injured by an AI system are now left empty-handed? Far from it: Dutch law already has similar instruments, as our colleague Reinoud Westerdijk and competing colleague Polo van der Putt have written in the recent past.[4]
i. Publication of evidence; duty to produce exhibits
Section 194 of the Dutch Code of Civil Procedure (the ‘former’ 843a CCP that applied until 1 January 2025) contains the duty to produce exhibits. By relying on this Section it is possible to claim inspection or a copy of documents the opposite party holds if the claimant has a legitimate interest therein, e.g. for the purpose of collecting evidence. However, Section 194 CCP seems to curb a successful reliance on this provision more than the Directive does. In either case, general ‘fishing expeditions’ are prevented, but under Section 194 CCP the claimant has to demonstrate a specific interest in the surrender of the documents claimed and has to specify the documents to be surrendered sufficiently. On this point, the AI Liability Directive would have been a welcome addition for injured parties.
ii. Presumption of causality; rule of reversal
Dutch law has the legal concept of the evidentiary presumption and Dutch courts apply the ‘rule of reversal’. If a culpable act has given rise to a risk of loss, and this risk has materialized later on, the causal link between this act and the loss arisen is in principle a given fact. It is up to the party that has allegedly caused the loss (in this context, the deployer or provider of the AI system) to prove otherwise. This rule of reversal is already applied where a loss is caused by defective or unsafe products that do not conform to statutory regulations. Westerdijk and Van der Putt concluded that the AI Liability Directive appears to be more conservative on this point, because it contains the additional requirement that it has to be made plausible that the violation of the standard can reasonably be assumed to have affected the functioning of the AI system. Such a requirement does not follow from the current Dutch rule of reversal. The party alleging to have incurred a loss by an AI system would be better off if the rule of reversal developed in Dutch case law were applied.
Conclusion
At present, we will have to wait and see whether any harmonised rules on liability and AI will come from Europe. If these rules do not or not timely make an appearance, the national regime will be fallen back on, which offers similar legal instruments for the party incurring a loss to attain or attempt to attain justice.
------------------
[1] https://commission.europa.eu/strategy-and-policy/strategy-documents/commission-work-programme/commission-work-programme-2025_en
[2] Apart from the European Commission, these are the European Parliament and the European Council.
[3] There has to be negligent conduct by this party – e.g. a fault caused by a human act or omission – that does not meet a duty of care under EU law or national law that offers direct protection against the damage.
[4] mr. P.G. van der Putt and mr. R.J.J. Westerdijk, AI en aansprakelijkheid: stappen naar een nieuw regime [AI and Liability: Steps Towards a New Regime] Tijdschrift voor Internetrecht, 2023 no. 1. This can also be consulted here: https://kvdl.com/artikelen/ai-en-aansprakelijkheid-stappen-naar-een-nieuw-regime
