Photo credit: Cryteria, via Wikimedia commons
1. Objectives of the proposal
On 28 September 2022, the European Commission published a proposal for an AI Liability Directive and an accompanying update of a complementary Product Liability Directive. In the preceding Report on Artificial Intelligence Liability, the Commission acknowledged the immense potential of artificial intelligence (AI). However, it has also identified the risks associated with it. For instance, connectivity of an AI-encompassing product can compromise its safety for users as it may be susceptible to cyber-attacks. Moreover, the outcomes of AI cannot always be predicted. To this end, ex ante risk assessments can be insufficient to address the possible wrongs. The opacity inherent in advanced AI-based products and systems makes it difficult to ascertain the responsibility of AI systems’ behaviours and choices. It is pivotal that humans can be able to understand how algorithmic decisions were reached in order to make a liability claim. Particularly, the opacity of AI systems can hinder victims in proving fault and causality in such cases. Consequently, the AI Liability Directive aims to ensure the provision of protection for victims of AI commensurate with those where damage has been caused by other products. It aims to increase trust in new technologies as well as to contribute to the ‘rollout of AI’ and improve its development in the internal market by preventing fragmentation and increasing legal certainty through harmonisation. Once adopted, these proposals will complement other AI regulation (e.g. the proposed AI Act) and establish liability rules for software and AI systems in the EU.
2. The proposed AI Liability Directive: scope
Other than what the name might suggest, the proposed AI Liability Directive does not provide any new ground of civil liability. That remains a matter of the national legislature, except when it comes to liability for defective products under the regime of the Product Liability Directive. Instead, the proposed AI Liability Directive provides for burden of proof rules on disclosure of evidence and rebuttable presumptions on a causal link. The rules cannot be invoked in every tort law case: only cases on fault liability fall within the scope of the AI Liability Directive. These are cases where liability for damage caused by (the use of) an AI system are based on fault. Fault encompasses wrongful actions and omissions. Due to the characteristics of AI systems, it can be difficult or prohibitively expensive to prove fault. Consequently, those suffering damage caused by an AI system might not get compensated for damage suffered, whereas those suffering damage from a non-AI system would be able to get compensated as they do not incur the same difficulties in proving fault. The proposed AI Liability Directive would address this discrepancy by providing rules on:
‘(a) the disclosure of evidence on high-risk artificial intelligence (AI) systems to enable a claimant to substantiate a non-contractual fault-based civil law claim for damages;
(b) the burden of proof in the case of non-contractual fault-based civil law claims brought before national courts for damages caused by an AI system.’ (art. 1 Proposal)
The proposed AI Liability Directive does not apply to risk-based liability claims. However, the proposed new Product Liability Directive does provide similar rules on disclosure of evidence and the burden of proof (art. 8 and 9).
The scope of the applicability of the proposed AI Liability Directive is partially limited to a specific category of AI system: the high-risk AI systems. For the definition of a high-risk AI system, the AI Liability Directive refers to the proposed AI Act. The AI Act identifies and lays down rules as per levels of risk associated with AI systems – those that carry (1) unacceptable risk, (2) high risk, and (3) limited risk. The fourth category – that of systems that pose a minimal risk (e.g. spam filters) – although within the material scope are not subject to any concrete rules. The first category – (1) unacceptable risk – concerns AI systems that are a clear threat to the safety, livelihoods and rights of persons (e.g. manipulation and social scoring systems). The third category (those of limited risk), is subject to specific transparency obligations due to their nature (e.g. deep fakes). High-risk AI systems represent those which are embedded in products subject to third-party assessment under sectoral legislation, and those which are not components of products but are deemed to be high-risk when used in certain areas (e.g. transport, education, safety components etc.). Such systems are subject to a set of requirements (e.g. risk assessments, mitigation systems, data quality, logging, and technical documentation) before being placed on the market.
The rules on disclosure of evidence as laid down in art. 3 of the proposed AI Liability Directive only apply to these high-risk AI systems. The rules on the burden of proof, however, apply to claims relating to all AI systems. (art. 4).
3. The proposed AI Liability Directive: disclosures and presumptions
3.1 Rebuttable presumption of a causal link
Article 4 introduces a rebuttable presumption of a causal link in the case of fault. It allows the courts to presume the causal connection between the fault of the defendant and the AI output (or failure to produce it) under three cumulative conditions. Firstly, the fault needs to be established (either by an assuming court or a claimant) consisting of non-compliance with the duty of care under EU or national law. Secondly, it must be ‘considerably likely’ that the fault has influenced the output of an AI system. Finally, damage by an AI system needs to be demonstrated. Paragraphs (2) and (3) differentiate between providers and users of AI systems.
The causal link concerning a claim for damages caused by a high-risk AI system shall not be presumed if the defendant demonstrates that sufficient evidence and expertise is reasonably accessible for the claimant to prove this causal link (art. 4(4)). When the claim concerns an AI system that is not high-risk, the presumption of the causal link shall only be applied where the national court considers it excessively difficult for the claimant to prove the causal link (art. 4(5)). Moreover, the defendant can always rebut any presumption regarding the causal link (art. 4(6)).
3.2 The disclosure of evidence
Article 3 of the proposed AI Liability Directive establishes the conditions regarding the disclosure of evidence and introduces a rebuttable presumption of non-compliance. This applies to high-risk AI systems as defined in the AI Act.
Article 3(1) of the Directive allows a court to order the disclosure of relevant evidence about specific high-risk AI systems that are suspected to have caused damage. Recital (16) confirms that this requirement has been unaccounted for by the AI Act proposal. However, the disclosure provided for in the AI Liability Directive does not seem to be absolute. Rather, it seems to be subject to a certain proportionality assessment since disclosure is only allowed to the extent necessary for sustaining the liability claim. To do that, national courts ought to consider the legitimate interests of all parties. Particularly, this applies in relation to the preservation of trade secrets and confidential information. The explanation notes convey that the aim is to strike a balance between ‘the claimant’s rights and the need to ensure that such disclosure would be subject to safeguards to protect the legitimate interests of all parties concerned, such as trade secrets or confidential information’. In other words, the goal is to strike a balance between claimant’s rights and the need for safeguards imposed by the court to preserve trade secrets or confidential information. The court will presume that the defendant did not comply with the duty of care if they refuse to disclose the requested information. In this case, the defendant can remedy that and rebut that presumption by providing evidence.
Recital (20) confirms that national courts should have the power to take specific measures to ensure the confidentiality of trade secrets during and after the proceedings in a proportionate manner in respect of balancing interests. Such measures could include restricting access to documents containing trade secrets and access to hearings or documents and transcripts thereof to a limited number of people. However, the courts cannot decide on this without considering the need to ensure the right to an effective remedy, fair trial and potential harm that could occur.
It is commendable that the EU is taking steps to address the information asymmetry between AI systems’ developers and individuals harmed by their creations. The (prospect of the) realisation of liability and compensation can provide an important incentive to AI providers and users to ensure the safe and correct functioning of their systems. Together with the Product Liability Directive, the proposed AI Act and other product safety rules such as the General Product Safety Directive, the European Commission is designing a comprehensive framework addressing the safety of AI systems and liability for damage caused by those systems.
Nevertheless, the AI Liability Directive harbours an important flaw that might have been overlooked: the AI Liability Directive offers defendants a way to avoid having to disclose evidence. As the proposal currently stands, if the defendant refuses to provide trade secret information about an AI system as evidence, they will be presumed as non-compliant with the duty of care. A defendant might decide it is strategically wiser to simply pay compensation in exchange for non-disclosure. After all, trade secrets are of great economic importance to such enterprises and one of the conditions for legal protection of trade secrets requires continuous efforts to keep information secret. In other words, non-compliance becomes a choice in order to avoid disclosure.
This is at odds with the drive for transparency of high-risk AI systems in the EU AI Act (art. 13). By offering an option to avoid transparency, the AI Liability Directive undermines this requirement for transparency indicated in the AI Act. This creates tension between these two new instruments. The EC could have chosen a clearer stance on transparency and its necessity, by carrying the requirement of transparency from the AI Act through to the AI Liability Directive.
There is an additional disadvantage to the route the EC has chosen. By avoiding the disclosure of the information necessary to establish fault in liability claims, one can avoid any flaws of the AI systems to be disclosed. This might take away any motivation to improve an AI system, as sufficient financial means offer the possibility to keep any shortcomings of the AI system hidden from the public eye. The lack of transparency could thereby lead to disincentivising the development and improvement of AI systems. Ultimately, this might negatively impact innovation and trust in AI.