BUSINESS MANAGEMENT GREAT WORK, ON TIME, NO PLAGARISM, A+ WORK

Artificial Intelligence and Tort Law: A ‘Multi-faceted’ Reality

Orian DHEU * & Jan DE BRUYNE

**

Abstract: Artificial Intelligence (AI) is becoming increasingly prevalent in our daily lives. Although AI systems bring many benefits, several legal and regulatory challenges remain as well. A field that has already attracted much attention is extra-contractual and product liability for damage involving AI systems. Accidents involving self-driving vehicles or surgical robots surely are a reason why tort and product liability are increasingly being discussed. Another reason relates to the intrinsic characteristics of AI such as opaqueness, autonomy, connectivity, data dependency or self-learning abil- ities, which make it difficult to trace back potentially problematic decisions made with the involvement of such systems. The attention for this academic field will only increase following the recent proposal by the European Commission (EC) regarding new liability rules for AI. Instead of focusing on one specific element, giving a rather general overview of the impact of AI on tort law or discussing the new rules, this article will take a more conceptual perspective and show that each policy decision regarding tort liability for damage involving AI actually entails several additional choices to be made. We will illustrate this with two use cases, namely the allocation of the burden of proof in an AI-context on the one hand and the adoption of strict liability regimes on the other hand. This starting point is important as normative recommendations and proposals risk to be one-layered and, consequently, too ‘simple’ or not realistic to directly imple- ment. Our research emphasizes the ‘multi-faceted’ reality of any proposal regarding the adoption of a new liability regime or modification to existing rules.

Résumé: L’intelligence artificielle dévient de plus en plus présente dans notre vie quotidienne. Cependant, malgré ses nombreux avantages, différents défis juridiques et réglementaires se dressent en chemin. Un domaine ayant déjà attiré beaucoup d’atten- tion touche à la responsabilité extra contractuelle et la responsabilité du fait des produits défectueux en lien avec un dommage impliquant un système d’intelligence artificielle. Les accidents impliquant des véhicules sans conducteur or des robots chirurgicaux sont assurément une raison pour laquelle de telles responsabilités sont actuellement discutées. Une autre raison à trait aux caractéristiques même de l’intelli- gence artificielle, à savoir son opacité, son autonomie, sa connectivité, sa dépendance aux données ou ses capacités d’auto apprentissage. De telles caractéristiques rendent beaucoup plus difficile le traçage de décisions potentiellement problématiques réalisées avec ces systèmes d’intelligence artificielle. L’attention portée à ce domaine de recher- che ne fera qu’accroitre au vu des récentes propositions de la Commission Européenne concernant de nouvelles règles pour l’intelligence artificielle. Au lieu de se concentrer sur un élément en particulier, donnant un aperçu général de l’impact de l’intelligence

* Doctoral researcher KU Leuven Centre for IT & IP Law – imec. ** Research expert AI and (tort) law KU Leuven Centre for IT & IP Law – imec; Assistant professor

eLaw Leiden. The article was finalized prior to the adoption of the rules on AI liability issued by the European Commission. Nevertheless, they have been integrated where relevant. Email: [email protected].

European Review of Private Law 2 & 3-2023 [261–298] © 2023 Kluwer Law International BV, The Netherlands.

261

artificielle sur le droit de la responsabilité ou de discuter de ces nouvelles règles, cet article adoptera une approche plus conceptuelle et s’efforcera de montrer que chaque décision politique sur ce sujet implique que plusieurs autres choix soient formulés. Nous illustrerons nos propos à travers deux exemples: d’un côté, la répartition de la charge de la preuve dans un contexte d’intelligence artificielle; et d’un autre côté, l’adoption d’un régime de responsabilité stricte. Ce point de départ est important compte tenu du fait que les recommandations et propositions normatives risquent d’être limités à un seul niveau, et de ce fait trop simples et pas assez réalistes afin d’être directement mises en œuvre. Notre recherche insiste sur le fait que toute proposition d’adoption d’un nou- veau régime de responsabilité implique de prendre en compte une réalité à plusieurs facettes.

Zusammenfassung: Künstliche Intelligenz (KI) durchdringt immer mehr alle Aspekte in unserem täglichen Leben. Obwohl KI-Systeme viele Vorteile bringen, sind diese an die Bewältigung einiger rechtlicher und regulatorischer Herausforderungen geknüpft. Dies hat bereits für Aufsehen gesorgt in dem Bereich der außervertraglichen Produkthaftung für Schäden an KI-Systemen. Unfälle mit selbstfahrenden Fahrzeugen oder chirurgischen Robotern sind sicherlich ein Grund, warum sowohl Schadensersatz als auch Produkthaftung zunehmend diskutiert werden. Ein weiterer Grund liegt in den intrinsischen Eigenschaften von KI, wie zum Beispiel Intransparenz, Autonomie, Konnektivität, Datenabhängigkeit oder Selbstlernfähigkeit, die es schwierig machen, potenziell problematische Entscheidungen unter Einbeziehung solcher Systeme zurückzuverfolgen. Die Aufmerksamkeit für diesen Forschungsbereich wird mit dem jüngsten Vorschlag der Europäischen Kommission (EK) zu neuen Haftungsregeln für KI noch zunehmen. Anstatt sich auf ein bestimmtes Element des Vorschlags zu konzen- trieren, einen eher allgemeinen Überblick über die Auswirkungen von KI auf das Deliktsrecht zu geben oder die neuen Regeln zu besprechen, wird dieser Artikel eine mehr konzeptionelle Perspektive einnehmen und zeigen, dass jede politische Entscheidung in Bezug auf die deliktische Haftung für Schäden mit zusätzliche Entscheidungen verbunden ist. Wir veranschaulichen dies anhand von zwei Anwendungsfällen, nämlich der Beweislastverteilung im KI-Kontext einerseits und der Einführung von Gefährdungshaftungsregelungen andererseits. Dieser Ausgangspunkt ist wichtig, da normative Empfehlungen und Vorschläge das Risiko beherbergen, dass sie „eindimensional“ und folglich zu „einfach“ oder nicht realistisch sind, um sie direkt umzusetzen. Unsere Forschung bestätigt die „vielschichtige“ Realität jedes Vorschlags zur Annahme eines neuen Haftungssystems oder zur Änderung bestehender Vorschriften.

1. Introduction

1. Artificial intelligence (AI) is becoming increasingly prevalent in our daily lives. AI systems are already used for a variety of purposes and deployed in many sectors. Some examples include self-driving vehicles, surgical robots, chatbots or virtual assistants.1 The rise of AI systems is no surprise considering their many benefits. They can be more accurate and efficient because they process information faster

1 See for an overview OECD, Artificial Intelligence in Society (OECD Publishing 2019), p 152.

262

and ‘better’ than humans.2 At the same time, however, several challenges exist as well. The reliance on AI may have an impact on the job market3 as well as entail the risk of reinforcing existing discriminatory practices (cf. bias).4 Other ethical ques- tions relate to the role humans may still play in an AI-era.5 More importantly, and as shown by the many contributions in this special issue, the commercialization of AI will also pose several legal and regulatory challenges.6

2. A field that has attracted much attention is tort and product liability for damage involving AI systems. Both academic research7 and policy initiatives8

have already addressed many pressing issues in this legal domain, such as the question under which circumstances AI is considered defective9 or whether soft- ware is a product falling within the product liability regime.10 Recently, the European Commission (EC) adopted two Proposals containing liability rules for

2 S.G. TZAFESTAS, Roboethics: A Navigating Overview (Athens: Springer 2015), p 147. 3 See e.g., C.B. FREY & M.A. OSBORN, ‘The future of employment: How susceptible are jobs to

computerisation?’, 14. Technological Forecasting and Social Change 2017, pp 254–280, doi: 10. 1016/j.techfore.2016.08.019.

4 See e.g., F. ZUIDERVEEN BORGESIUS, Discrimination, artificial intelligence, and algorithmic decision- making (Council of Europe 2018), p 48.

5 See e.g., H. FRY, Hello World: How to be Human in the Age of the Machine (New York: Random House 2018), p 320.

6 See e.g., M. EBERS & S. NAVAS (eds), Algorithms and Law (Cambridge: Cambridge University Press 2020), p 319; B. CUSTERS & E. FOSCH-VILLARONGA, Law and Artificial Intelligence Regulating AI and Applying AI in Legal Practice (Berlin: Springer 2022), p 569; J. DE BRUYNE & C. VANLEENHOVE, Artificial Intelligence and the Law (Antwerpen: Intersentia 2021), p 520.

7 See e.g., the many contributions in S. LOHSSE et al. (eds), Liability for Artificial Intelligence and the Internet of Things (Baden-Baden: Nomos 2019), p 352; M. DIAMANTIS, ‘Vicarious Liability for AI’, U Iowa Legal Studies Research Paper No. 2021-27; J. DE BRUYNE, E. VAN GOOL & T. GILS, ‘Tort Law and Damage Caused by AI Systems’, in J. DE BRUYNE & C. VANLEENHOVE (eds), Artificial Intelligence and the Law, pp 259–403 with many references.

8 See e.g., EUROPEAN PARLIAMENT, ‘Report with recommendations to the Commission on Civil Law Rules on Robotics’ (2017) 2015/2103(INL); EUROPEAN PARLIAMENT, ‘Report with recommendations to the Commission on a civil liability regime for artificial intelligence’ (Report) 2020/2014(INL); EXPERT GROUP ON LIABILITY AND NEW TECHNOLOGIES – NEW TECHNOLOGIES FORMATION, ‘Liability for Artificial Intelligence and other emerging digital technologies’ (Nov. 2019), p 70 (further referred to as EXPERT REPORT 2019); EUROPEAN COMMISSION, ‘Report on the safety and liability implications of Artificial Intelligence, the Internet of Things and robotics’ (Feb. 2020), p 18 (further referred to as EUROPEAN COMMISSION, ‘Report on Safety and Liability’).

9 See in particular J.S. BORGHETTI, ‘How can Artificial Intelligence be Defective?’, in S. LOHSSE et al. (eds), Liability for Artificial Intelligence and the Internet of Things, pp 63–76.

10 See e.g., D. FAIRGRIEVE & E. RAJNERI, ‘Is Software a Product under the Product Liability Directive?’, 1. Zeitschrift für Internationales Wirtschaftsrecht 2019, pp 24–27; B. KOCH et al., ‘Response of the European Law Institute to the Public Consultation on Civil Liability – Adapting Liability Rules to the Digital Age and Artificial Intelligence’, 13. Journal of European Tort Law 2022, pp 34–36. Giovanni Comandé even concludes that ‘[even] when duly adapted, traditional tort liability rules are ineffective in sorting the allocation of liability and costs related to AI’ (G. COMANDÉ,

263

AI and providing some guidance on many of these issues. One Proposal revises the Product Liability Directive (‘revised PLD’)11 and another one introduces extra-con- tractual civil liability rules for AI systems (‘AI Liability Directive’).12 Accidents involving self-driving vehicles13 or surgical robots14 surely are a reason why tort and product liability are increasingly being discussed. Another reason relates to the intrinsic characteristics of AI such as opaqueness, autonomy, connectivity, data dependency or self-learning abilities, which make it difficult to trace back potentially problematic decisions made with the involvement of such systems.15 This in turn makes it challenging for victims to claim compensation under the current European Union (EU) and national (fault-based) liability regimes.16 While it will not be dis- cussed in this article, another crucial aspect resides in defining the concept of AI as such, should that be even possible. Indeed, any legal intervention presupposes clearly identifying the scope of application of the law. However, when contemplating legisla- tion aimed at AI applications, defining AI itself will be challenging as it is an ‘umbrella term’.17

3. Instead of focusing on one specific element, giving a rather general overview of the impact of AI on tort law (as so many contributions do) or discussing the new AI liability rules proposed by the EC,18 this article will take a more conceptual

‘Multilayered (Accountable) Liability for Artificial Intelligence’, in S. LOHSSE et al. (eds), Liability for Artificial Intelligence and the Internet of Things, p 176).

11 Proposal for a Directive of the European Parliament and of the Council on liability for defective products, European Commission, COM(2022) 495 final, 2022 (‘revised PLD’).

12 Proposal for a Directive of the European Parliament and of the Council on adapting non-contrac- tual civil liability rules to artificial intelligence, European Commission, COM(2022) 496 final, 2022 (‘AI Liability Directive’).

13 For example Tesla’s Blog, ‘A Tragic Loss’ (30 Jun. 2016), https://www.teslamotors.com/blog/ tragic-loss.

14 Mracek v. Bryn Mawr Hospital, 363 Fed. Appx. 925 (3d Cir. 2010) as reported in T.N. WHITE & S.D. BAUM, ‘Liability for Present and Future Robotics Technology’, in P. LIN et al. (eds), Robot Ethics 2.0: From Autonomous Cars to Artificial Intelligence (New York: Oxford University Press 2017), pp 66–79.

15 See e.g., EXPERT REPORT 2019, pp 32–34; B. KOCH et al., 13. Journal of European Tort Law 2022, pp 50–56; Recitals (3)-(7) AI Liability Directive.

16 EUROPEAN COMMISSION, ‘White Paper on Artificial Intelligence – A European approach to excellence and trust’, 19 Feb. 2020, COM(2020) 65 final, p 13.

17 See e.g., A. BERTOLINI, ‘Artificial Intelligence does not exist: defying the technology-neutrality narrative in the regulation of civil liability for advanced technologies’, 2. Europa E Diritto Privato 2022, pp 369–371. We use the concept as it is defined in Art. 3(1) in the Proposal of a Regulation on AI. The definition has, however, been criticized (e.g., M. EBERS et al., ‘The European Commission’s Proposal for an Artificial Intelligence Act – A Critical Assessment by Members of the Robotics and AI Law Society (RAILS)’, 4. J – Multidisciplinary Scientific Journal 2021, pp 590–591).

18 See for an extensive analysis: O. DHEU et al., ‘The European Commission’s Approach to Extra- Contractual Liability and AI – A First Analysis and Evaluation of the Two Proposals’, KU Leuven Centre for IT & IP (CiTiP) Working Paper 2022 (6 Oct. 2022), https://papers.ssrn.com/sol3/ papers.cfm?abstract_id=4239792.

264

perspective and show that each policy decision regarding tort liability for damage involving AI actually entails several additional choices to be made. We will illus- trate this with two use cases, namely the allocation of the burden of proof in an AI- context on the one hand (part 2) and the adoption of strict liability regimes on the other hand (part 3). This starting point is important as normative recommenda- tions and proposals risk to be one-layered and, consequently, too ‘simple’ or not realistic to directly implement. Our research emphasizes the ‘multi-faceted’ reality of any proposal regarding the adoption of a new liability regime or mod- ification to existing rules.19 The most important findings are summarized in a conclusion (part 4).

2. The Allocation of the Burden of Proof in an AI-Context

4. Whereas substantive provisions regarding AI and tort law have already been assessed by many scholars, procedural elements have not (yet) attracted the same amount of attention. Nevertheless, the New Technologies Formation concludes in its report from 2019 that the ‘application of liability frameworks [in the field of digital technologies] in practice is also affected by challenges in the field of procedural law’.20

That is because the law of evidence can serve as an instrument to attain a desired solution.21 Existing liability rules were framed decades ago based on even older concepts and incorporating a primarily ‘anthropocentric and monocausal model’ of causing damage.22 The adequacy of these existing liability rules may thus be question- able when applied to AI.23 Moreover, several parties are also involved in the supply chain of AI systems. One can think of the developers of the software/algorithm, the producer of the hardware, owners/keepers of the AI product, suppliers of data, public authorities, network infrastructure managers or the users of the product.24 This may lead to uncertainty as to whom will eventually be held liable, and especially who to target when damage is incurred. Persons having suffered harm involving AI systems may not have adequate access to the information and, therefore, evidence in order to prove their case in court. They may indeed have less successful claims compared to damages involving ‘traditional’ products.25 Litigation could become burdensome and

19 See on the concept of a multi-layered accountable liability for AI also G. COMANDÉ, in Liability for Artificial Intelligence and the Internet of Things, pp 165–183.

20 EXPERT REPORT 2019, p 29. 21 I. GIESEN, ‘The Reversal of the Burden of Proof in the Principles of European Tort Law. A

Comparison with Dutch Tort Law and Civil Procedure Rules’, 6. Utrecht Law Review 2010, p 30. 22 EXPERT REPORT 2019, p 19. 23 EXPERT REPORT 2019, p 19. 24 See COUNCIL OF EUROPE, EXPERT COMMITTEE ON HUMAN RIGHTS DIMENSIONS OF AUTOMATED DATA PROCESSING

AND DIFFERENT FORMS OF ARTIFICIAL INTELLIGENCE (MSI-AUT), ‘Responsibility and AI’, 2019, DGI (2019), pp 05, 11, at 62–64 referring to the ‘many hands’ problem.

25 EUROPEAN COMMISSION, ‘White Paper on Artificial Intelligence’, p 13.

265

expensive for victims, leaving them without ‘effective access to justice’.26 It is, however, important that victims of accidents involving AI are not confronted with a lower level of protection compared to other (traditional) products and/or services for which they would get compensation under national law. Otherwise, the societal acceptance of those AI systems could be hampered, resulting in hesitance to use them.27

5. Against this background, the allocation of the burden of proof can be a strategic tool for policymakers and courts. It allows relevant actors to create incentives to achieve desired out-of-court behaviour,28 which may be relevant in an AI context. In this section, we first give a general overview of the challenges posed by AI to some ‘traditional’ procedural rules (part 2.1.). We then focus on the reversal of the burden of proof or the adoption of rebuttable presumptions (part 2.2.). Finally, we discuss some modalities that need to be considered in (potential) legislation that includes provisions on the allocation of the burden of proof (part 2.3.). This analysis is relevant considering that the recent proposals by the EC on liability rules for AI will likely be the subject of debate in the Council as well as in the European Parliament (EP).

2.1. General Background on Procedural Law

6. A ‘general, worldwide accepted rule’29 in the law of evidence is that each party has to prove its claims and contentions (actori incumbit probatio).30 The applica- tion of this procedural rule can, however, be challenging when accidents involve AI systems as such technologies can be complex and inaccessible. As explained in the Expert Group Report, some AI algorithms are not easily understandable and inter- pretable but come in forms of ‘black boxes’ that evolve through self-learning. Victims are, therefore, confronted with the increasingly daunting task of trying to identify and prove AI systems as their source of harm.31 Moreover, the injured party will likely be a natural person, while the defendant will mostly be a legal person with considerable knowledge on the specific AI system or technology in

26 EXPERT REPORT 2019, p 35. 27 EUROPEAN COMMISSION, ‘Report on Safety and Liability’, p 13; AD HOC COMMITTEE ON ARTIFICIAL

INTELLIGENCE – CAHAI, ‘Feasibility Study’, CAHAI(2020)23, 17 Dec. 2020, p 38, no. 113. Also see EXPERT REPORT 2019, p 35.

28 F. GOMEZ, ‘Burden of Proof and Strict Liability: An Economic Analysis of a Misconception’, InDret, Barcelona (Jan. 2001), p 7, https://indret.com/wp-content/themes/indret/pdf/040_en.pdf/.

29 I. GIESEN, ‘The Burden of Proof and other Procedural Devices in Tort Law’, in H. KOZIOL et al. (eds), European Tort Law 2008 (Wien: Springer 2009), p 50.

30 M. KAZAZI, Burden of Proof and Related Issues: A Study on Evidence Before International Tribunals (Gravenhage: Martinus Nijhoff Publishers 1996), p 378. See e.g., Art. 8.4, para. 1 Belgian Civil Code (Act of 13 Apr. 2019, BS 14 May 2019, p 46353); Art. 870 Belgian Judicial Code.

31 EXPERT REPORT 2019, p 33; Recitals (3)-(7) AI Liability Directive.

266

general. Most parties will thus not have access to the necessary information to build a case in court.32

7. Under the PLD, the burden of proof is high as well. A victim has to prove that the product caused the damage because it is defective, implying that it did not provide the safety one is legitimately entitled to expect.33 Already in 2006, the ECmentioned in its report on the application of the PLD that the difficulty in establishing the defect:

mainly arises from perceived difficulties in proving claims due to a lack of legal or other resources needed to investigate themproperly, or to an inability to gain access to essential information. Such problems are seen to be particularly acute in relation to highly technical products, or where the alleged injuries are of a complicated nature.34

De Bruin argues that proving the defect in a software demands a very thorough understanding of its (mal)functioning. Proving the defect will, therefore, be challenging.35 Likewise, the 2019 Expert Group Report concludes that proving that some hardware defect caused damage to a person is difficult. It is even more challenging to establish that the underlying cause was a flawed algorithm. Things become even more complex when the algorithm that may have caused harm has evolved over time through its self-learning abilities enabled by machine learning and/or deep learning techniques, while being fuelled by external data that it has collected since the start of its operation.36 It is also uncertain what exactly con- stitutes a defect of an advanced AI system. For instance, if an AI diagnosis tool delivers a wrong diagnosis, ‘there is no obvious malfunctioning that could be the basis for a presumption that the algorithm was defective’.37 It may thus be difficult

32 EUROPEAN COMMISSION, ‘White Paper on Artificial Intelligence’, p 13. See also D. WUYTS, ‘The Product Liability Directive – More than Two Decades of Defective Products in Europe’, Journal of European Tort Law 2014 (5), p 24.

33 Council Directive 85/374/EEC of 25 Jul. 1985 on the approximation of the laws, regulations and administrative provisions of the Member States concerning liability for defective products, OJ L 210 (further referred to as the ‘PLD’).

34 Report from the Commission to the Council, the European Parliament and the European Economic and Social Committee – Third report on the application of Council Directive on the approximation of laws, regulations and administrative provisions of the Member States concerning liability for defective products (85/374/EEC of 25 Jul. 1985, amended by Directive 1999/34/EC of the European Parliament and of the Council of 10 May 1999, COM/2006/0496 final.

35 R. DE BRUIN, ‘Autonomous Intelligent Cars on the European Intersection of Liability and Privacy’, 7. European Journal of Risk Regulation 2016, p 491. See also D. WUYTS, 5. Journal of European Tort Law 2014, p 24.

36 EXPERT REPORT 2019, p 20. 37 J.-S. BORGHETTI, in Liability for Artificial Intelligence and the Internet of Things, p 67 as referred to

in M. BUITEN et al., ‘EU Liability Rules for the Age of Artificial Intelligence’, Centre on Regulation in Europe (9 Apr. 2021), pp 34–35.

267

and costly for consumers to prove the defect when they have no expertise in the field, especially when the computer program is complex and not readable ex post.38

Moreover, depending on the Member State where the action is introduced, the requirement of proof of defect may be different.39 This may thus place some consumers in a more difficult situation than the ones in a neighbouring country.40

Proving causation is not straightforward either for victims under the PLD,41

and by extension other similar national (strict) liability regimes, such as the liability of the custodian of a defective good under Article 1384, paragraph 1 of the Old Belgian Civil Code (‘OBCC’).42 Proving causality in the context of AI harm may be difficult, especially if some human supervision is still required. The injured party may have difficulty showing that the AI system, and thus not his/her negligence, caused the harm. As Buiten and others rightly note, the ‘assessment of the causal link will often require expert advice, the cost of which may discourage injured parties from suing’.43 Moreover, many actors may be involved in the production, supply and use of goods incorporating software or of AI systems. The Expert Group explains that the complex digital eco-system and the multiplicity of actors makes it very difficult to determine who may be held liable for the damage caused to the victim(s).44 In sum, the damage in an AI-context may result from ‘a conjunction of intertwined effective and successive causes that have been collectively triggered by multiple actors’.45

38 Also see Recitals (30)-(31) revised PLD (‘Injured persons, are, however, often at a significant disadvantage compared to manufacturers in terms of access to, and understanding of, information on how a product was produced and how it operates. This asymmetry of information can undermine the fair apportionment of risk, in particular in cases involving technical or scientific complexity’).

39 See e.g., D. FAIRGRIEVE et al., ‘The Product Liability Directive: Time to Get Soft’, 4. Journal of European Tort Law 2013, pp 5–9. See extensively C. VAN DAM, European Tort Law (Oxford: Oxford University Press 2013), p 654.

40 C. DE MEEUS, ‘The Product Liability Directive at the Age of the Digital Industrial Revolution: Fit for Innovation?’, 4. Journal of European Consumer and Market Law 2019, p 152 referring to S. BECK, ‘The problem of ascribing legal responsibility in the case of robotics’, AI & Society 2016, p 474.

41 See extensively M. MARTIN-CASALS, ‘Causation and Scope of Liability in the Internet of Things’, in S. LOHSSE et al. (eds), Liability for Artificial Intelligence and the Internet of Things, pp 201–228.

42 See for an analysis of causation in the context of autonomous vehicles in Belgium M. KRUITHOF & T. VERHEYEN, ‘Toerekening van verkeersongevallen aan autonome motorvoertuigen’, in J. DE BRUYNE

(ed.), Autonome motorvoertuigen. Een multidisciplinair onderzoek naar de maatschappelijke impact (Brugge: Vanden Broele 2021), pp 239–270.

43 M. BUITEN et al., ‘EU Liability Rules for the Age of Artificial Intelligence’ (9 Apr. 2021), p 55 with further references.

44 EXPERT REPORT 2019, p 33; J. DE BRUYNE et al., in Artificial Intelligence and the Law, p 398. Also see supra n. 4.

45 B. KOCH et al., 13. Journal of European Tort Law 2022, p 52. Also see EUROPEAN LAW INSTITUTE, ‘Guiding Principles for Updating the Product Liability Directive for the Digital Age’, ELI Innovation Paper Series (21 Jan. 2021), p 9.

268

8. An additional hurdle is that the elements of a claim in tort law are governed by national law. An example is the requirement of causation including procedural questions such as the standard of proof or the laws and practice of evidence.46 The application of national rules can result in diverse outcomes, which may not always be desired in an AI context. This can be illustrated with an example of the situation in Belgium. The traditionally accepted theory in Belgium is the doctrine of the equivalence of conditions.47 Under this rule, courts are required to accept legal causality between a fact and the incurred damage if it has been established that the invoked fact was necessary in the given circumstances for the damage to occur (the conditio sine qua non or ‘but-for test’). As Kruithof explains, a ‘fact that in concreto was necessary for the harm to occur is a cause even if a later intervening fact was also necessary’.48 Moreover, a fact that was necessary for the damage to occur the way it did is a cause even if this damage was not a necessary/normal or foreseeable consequence of the fact.49 The doctrine of the equivalence of conditions can lead to persons being held liable in a seemingly unfair way for damage they could not have reasonably foreseen. For instance, in an AI-context, a causal link can be established between the harm that is caused by a robot to a patient during a surgical procedure and the original quarrel between neighbours causing that person to end up in hospital in the first place. The neighbour may be held liable as his/her behaviour was a necessary factual condition for the damage that eventually occurred. Belgian judges, however, do not always strictly apply this doctrine.50 The unpredictability of damage within long causal chains may sometimes lead to an exemption from liability.51 It is worth noting that the proposed Article 5:162 that may be included in the new Belgian Civil Code would continue to start from the conditio sine qua non test. Nevertheless, it would

46 B. KOCH et al., 13. Journal of European Tort Law 2022, pp 44–46 and 57–58. Similarity in the context of the PLD: D. WUYTS, 5. Journal of European Tort Law 2014, pp 23–24.

47 See e.g., T. VANSWEEVELT en B. WEYTS, Handboek buitencontractueel aansprakelijkheidsrecht (Antwerpen: Intersentia 2009), pp 763–873 with further references; J.L. FAGNART, ‘Petite naviga- tion dans les méandres de la causalité’, RGAR 2006, 14.080, no. 3.

48 M. KRUITHOF, Tort Law in Belgium (Alphen aan den Rijn: Kluwer Law International 2018), p 126 with references to case law.

49 M. KRUITHOF, Tort Law in Belgium, p 126 with references to case law. 50 J. DE BRUYNE et al., in Artificial Intelligence and the Law, p 398. Also see M. KRUITHOF, ‘Oorzaak of

aanleiding? Geen causaal verband zonder causale bijdrage’, in T. VANSWEEVELT & B. WEYTS (eds), Actuele ontwikkelingen in het aansprakelijkheidsrecht en verzekeringsrecht. Iste Interuniversitair Congres over Aansprakelijkheids – en Verzekeringsrecht (Antwerpen: Intersentia 2015), pp 139– 208 concluding that another rule can better describe Belgian positive law on causation in civil liability. Under this rule a fact is a cause of a loss if (1) the loss would in the given circumstances not have occurred as it specifically occurred without the fact and (2) the fact has increased the specific risk of which the specific loss occurrence was a realization.

51 See the many references in M. KRUITHOF, Actuele ontwikkelingen in het aansprakelijkheidsrecht en verzekeringsrecht 2015, pp 139–208; H. BOCKEN & I. BOONE, ‘Causaliteit in het Belgische recht’, TPR 2002, pp 1646–1670.

269

contain an exception for situations in which the relationship between the fact leading to the damage and the suffered damage is so remote that it would be ‘manifestly unreasonable’ to impose liability on the defendant. It remains to be seen how this would be applied in practice to cases where damage is caused by AI systems.52

2.2. Reversing the Burden of Proof or Working With Rebuttable Presumptions

9. Against this background, the idea of reversing the burden of proof or working with rebuttable presumptions for accidents involving AI becomes appealing.53 A reversal of the burden of proof may lead to a ‘tightening up of liability’,54 which, according to law and economics scholars, may enhance the deterrent function of tort law and hence result in safer behaviour.55 According to such scholars, the purpose of damage payments in tort law is not only to compensate injured parties but to provide incentives for potential injurers to take efficient cost-justified pre- cautions to avoid causing the accident.56 The fact that someone can be held liable ex post can provide the necessary incentives ex ante to act in such a way to prevent liability.57 Law and economics scholars thus argue that injurers might adopt cost- justified safety measures if the system holds them liable for the injury costs they generate.58 In an AI context, the reversal of the burden of proof or working with a rebuttable presumption could be considered for several reasons.59

52 Article 5.162, para. 2, Avant-projet de loi portant insertion des dispositions relatives à la responsabilité extracontractuelle dans le nouveau Code civil 1 Sep. 2019.

53 One can also work with rebuttable presumptions (C. VOLPIN, ‘The Ball is in Your Court: Evidential Burden of Proof and Proof-Proximity Principle in EU Competition Law’, 51. CMLR (Common Market Law Review) 2014, p 1165; C. CAUFFMAN, ‘Robo-liability: The European Union in search of the best way to deal with liability for damage caused by artificial intelligence’, 25. Maastricht Journal of European and Comparative Law 2018, p 531).

54 I. GIESEN, 6. Utrecht Law Review 2010, p 25. 55 See e.g., G. CALABRESI, The Costs of Accidents: A Legal and Economic Analysis (Yale: Yale University

Press 1970), p 340; R.A. POSNER, ‘A Theory of Negligence’, 1. Journal of Legal Studies 1972, p 29. 56 P.H. RUBIN, ‘Law and Economics’, in D.R. HENDERSON (ed), The Concise Encyclopedia Economics

(Liberty Fund 2008), www.econlib.org/library/Enc/LawandEconomics.html; M.G. FAURE et al., ‘Naar een Kostenoptimalisatie van de letselschaderegeling: een verkenning’, 4. Aansprakelijkheid, Verzekering & Schade 2011, online PDF version Kluwer Navigator 3–4.

57 M.G. FAURE & T. HARTLIEF, Nieuwe risico’s en vragen van aansprakelijkheid en verzekering (Alphen aan Den Rijn: Kluwer 2002), p 19; I. GIESEN, 75. Albany Law Review 2012, p 181.

58 S.D. SMITH, ‘Critics and the Crisis a Reassessment of Current Conceptions of Tort Law’, 72. Cornell Law Review1987, p 772with further references in footnote 28; T.C. JR.GALLIGAN, ‘Deterrence: TheLegitimate Function of the Public Tort’, 58. Washington and Lee Law Review 2001, p 1020; W.M. LANDES & R.A. POSNER, The Economic Structure of Tort Law (Harvard: Harvard University Press 1987), p 10.

59 See in general J. DE BRUYNE, Third-Party Certifiers (London: Kluwer Law International 2019), pp 263–286 and applied to AI J. DE BRUYNE & O. DHEU, ‘Liability for Damage Caused by Artificial

270

10. One could, for instance, burden the party whose ‘bad’ behaviour should be deterred. Burdening the developer of AI systems to prove compliance with its obliga- tions during the process can make litigation a worse outcome for this defendant due to the time and efforts put in the legal procedure. This may give that party a greater incentive to avoid litigation, which it can do by taking the necessary (safety) precau- tions during the development of an AI system.60 A more important and relevant, but somehow more contested, reason justifying a reversal of the burden of proof or a rebuttable presumption relates to a party’s possibility to gather evidence. The burden of proof could be placed on the party with better access to relevant information.61 One thus has to examine for which actor collecting the evidence would be the less burdensome.62 This can, for instance, be the developer of an AI system or the manufacturer of a product incorporating AI.63 More generally, a reversal of the burden of proof may be desirable when there is an information (and resource) asymmetry between the parties.64 The existence of an information asymmetry is also a reason why (legal) presumptions are created. Presumptions try to equalize the position of the parties in those cases where all the factual evidence is in the hands of the defendant.65 As already mentioned, AI developers or producers and companies marketing AI will in most cases possess more relevant information about the actual AI systems. Such parties may be required by law or following ‘best practices’ to already keep the necessary documents and track records that they obtained during the development process (cf. the relevant provi- sions in the draft AI Act66). Victims of damage caused by AI systems may sometimes be

Intelligence – Some Food for Thought and Future Debate’, in P. Morgan, Tort Liability and Autonomous Systems Accidents (Cheltenham: Edward Elgar 2023), forthcoming.

60 Compare C.W. SANCHIRICO, ‘A Primary Activity Approach to Proof Burdens’, 37. The Journal of Legal Studies 2008, p 276; J. DE BRUYNE, Third-Party Certifiers (2019), p 339.

61 See for an overview: C.W. SANCHIRICO, 37. The Journal of Legal Studies 2008, p 275, footnote 3; E. A. POSNER, ‘Fault in Contract Law’, 107. Michigan Law Review 2009, p 1444.

62 B. ALLEMEERSCH et al., ‘Overzicht van rechtspraak. Het burgerlijk bewijsrecht 2000–2013, [De burgerlijke bewijslast] Concrete toepassing van de basisregel’, 43. Tijdschrift voor Privaatrecht 2015, p 726; J. DE BRUYNE, Third-Party Certifiers, 2019, p 342.

63 J. DE MOT, ‘De verdeling van de bewijslast economisch bekeken’, in J. De Mot (ed), Liber amicorum Boudewijn Bouckaert. Vrank en vrij (Brugge: Die Keure 2012), p 24; M. HEMRAJ, Credit Rating Agencies: Self-regulation, Statutory Regulation and Case Law Regulation in the United States and European Union (Cham: SpringerLink 2015), p 175; J. DE BRUYNE, Third-Party Certifiers 2019, pp 342–343; Court of Appeal Antwerpen 15 Jun. 2015, DAOR (Le Droit des Affaires-Het Ondernemingsrecht) 2016, p 31.

64 M.G. FAURE, Tort Law and Economics (Cheltenham: Edward Elgar 2009), p 30; I. GIESEN, Bewijs en aansprakelijkheid: een rechtsvergelijkend onderzoek naar de bewijslast, de bewijsvoeringslast, het bewijsrisico en de bewijsrisico-omkering in het aansprakelijkheidsrecht (Den Haag: Boom Juridische Uitgevers 2001), p 537; J. DE BRUYNE, Third-Party Certifiers 2019, p 343.

65 C. VOLPIN, 51. CMLR 2014, p 1165. 66 EUROPEAN COMMISSION, ‘Proposal for a Regulation of the European Parliament and of the Council

laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) and amending

271

considered as weak or ‘unsophisticated’67 persons that trust AI developers or producers. A reversal of the burden of proof may thus be relevant in such circumstances.68

11. Indeed, several policy initiatives and proposals have already suggested to reverse the burden of proof or to work with rebuttable presumptions in certain circumstances. Although the New Technologies Formation Expert Group concludes that ‘[a]s a general rule, the victim should continue to be required to prove what caused her harm’,69 it also acknowledges that Member States may alleviate the burden of proof for victims of emerging digital technologies who have ‘a hard time proving that the technology in question was the actual cause of their harm’.70 It adds that ‘where a particular technology increases the difficulties of proving the existence of an element of liability beyond what can be reasonably expected, victims should be entitled to facilitation of proof’.71 It subsequently suggests several procedural elements to that end. The Expert Group proposes that where ‘the damage is of a kind that safety rules were meant to avoid, failure to comply with such safety rules, should lead to a reversal of the burden of proving (a) causation, and/or (b) fault, and/or (c) the existence of a defect’.72 It adds that if:

it is proven that an emerging digital technology caused harm, and liability there- fore is conditional upon a person’s intent or negligence, the burden of proving fault should be reversed if disproportionate difficulties and costs of establishing the relevant standard of care and of proving their violation justify it.73

The burden of proving causation could also be alleviated ‘in light of the challenges of emerging digital technologies if a balancing of the listed factors warrants doing so’74 (e.g., the likelihood that the technology at least contributed to the harm or the

certain Union legislative acts’, COM(2021) 206 final (with regard to record keeping obligations for high risk AI systems, see for instance Arts 12 and 20).

67 See in context of credit ratings Bathurst Regional Council v. Local Government Financial Services Pty Ltd (No 5) (2012) FCA 120, paras 2767–2778; ABN AMRO Bank NV v. Bathurst Regional Council (2014) FCAFC 65, paras 580, 599, 890–891, 1211 and 1263–1269.

68 I. GIESEN, Bewijs en aansprakelijkheid: een rechtsvergelijkend onderzoek naar de bewijslast, de bewijsvoeringslast, het bewijsrisico en de bewijsrisico-omkering in het aansprakelijkheidsrecht, 2001, pp 418 and 537–538. The idea of working with a reversal of the burden of proof has already been suggested in other contexts characterized by information asymmetry as well such as third- party certifiers (see extensively J. DE BRUYNE, Third-Party Certifiers 2019, p 446).

69 EXPERT REPORT 2019, p 49. 70 EXPERT REPORT 2019, pp 29–30. 71 EXPERT REPORT 2019, p 4. 72 EXPERT REPORT 2019, p 48. 73 EXPERT REPORT 2019, pp 8 and 52. 74 EXPERT REPORT 2019, p 49.

272

kind and degree of harm potentially and actually caused).75 More importantly, Article 4 of the AI Liability Directive introduces a rebuttable presumption of a causal link in the case of fault. The presumption of causality only applies when different conditions are met: (1) the proof of a defendant’s fault (consisting in the non-compliance with a duty of care), (2) it can be considered reasonably likely that the fault has influenced the (failure of) output produced by the AI system and (3) the incurrence of damage. The defendant, however, has the right to rebut the presumption of causality. Additional provisions apply for providers and users of high-risk AI systems violating certain provisions of the AI Act, as well as for low-risk AI systems and in case a defendant uses the AI system in the course of a personal, non-professional activity.76 The Resolution issued by the EP in October 2020 also contains a reversal of the burden of proof regarding fault-based liability for operators of low risk AI systems.77

12. Allocating the burden of proof has especially attracted attention in the con- text of product liability. The Expert Group report suggests that in the context of digital technologies (such as AI) that caused harm, the burden of proving the defect could be reversed ‘if there are if there are disproportionate difficulties or costs pertaining to establishing the relevant level of safety or proving that this level of safety has not been met’.78 This implies that the burden of proof with regard to the requirement of defect in Article 4 of the PLD would be turned around allowing to hold the manufacturer liable unless he/she is able to prove that the product was not defective.79 Some scholars have even advocated to reverse the burden of proof with regard to causation.80 It has also been noted that an additional challenge in the digital context is that a common feature of both products with digital elements and of IoT systems is a reliance on external data to determine how the product or system operates. Where such data is supplied from external sources, proving both defectiveness and a causal link with the injury or damage sustained becomes very difficult indeed. Externally supplied data could cause a product or system to malfunction. The European Law Institute (ELI)ELI Innovation Paper has argued that in the context of AI, the individual victim should not be burdened with having to exclude the relevance of external data in the harm’s occurrence. However, the

75 EXPERT REPORT 2019, pp 8 and 49–50. 76 See for an extensive analysis: O. DHEU et al., ‘The European Commission’s Approach to Extra-

Contractual Liability and AI – A First Analysis and Evaluation of the Two Proposals’ (6 Oct. 2022), pp 10–14.

77 Article 8 European Parliament resolution of 20 Oct. 2020 with recommendations to the Commission on a civil liability regime for artificial intelligence, 2020/2014(INL) (further referred to as ‘European Parliament Resolution 2020’). Also see supra n. 18.

78 EXPERT REPORT 2019, p 42. 79 G. WAGNER, ‘Robot Liability’, in S. LOHSSE et al. (eds), Liability for Artificial Intelligence and the

Internet of Things, p 47; C. DE MEEUS, 4. Journal of European Consumer and Market Law, p 152. 80 See e.g., R. DE BRUIN, 7. European Journal of Risk Regulation 2016 p 495.

273

producer could be burdened with an obligation to prove that it was not the product or systems that itself caused the injury or damage, but that it was the ‘externally supplied data’.81 Against this background, it is no surprise that the revised PLD also introduces presumptions of defectiveness and causality that apply under cer- tain conditions. For instance, a product’s defect is presumed when the claimant establishes that the damage was caused by an obvious malfunction of the product during normal use or under ordinary circumstance (Art. 9.2 (c)). The causal link between the defectiveness of the product and the damage can on the other hand be presumed where it has been established that the product is defective and the damage which is caused is of a kind typically consistent with the defect in question (Art. 9.3). A defendant can rebut any of the presumptions included in the revised PLD (Art. 9.5).82

2.3. Some Modalities to Be Considered in a Potential Legal Solution Regarding the Burden of Proof

13. Several EU initiatives have thus already relied on a reversal of the burden of proof or rebuttable presumptions. However, such (procedural) mechanisms should be carefully addressed. They may have an impact on innovation83 and could significantly change ‘the current distribution of risks to the detriment of the manufacturers’.84 It would also risk moving away from the established principles under the PLD.85 Moreover, as has been noted by Schütte, one of the risks with such an approach resides in the possibility for litigants to engage in ‘excessive or abusive litigation’ where the victims could channel claims against ‘all possible defendants who are then all using resources in litigation’ (whereas only one of them is the true cause of harm).86 As such, a reversal could foster a ‘claims culture’,87 as it would encourage ‘spurious claims’.88 Wagner rightly concludes in this regard that ‘lawmakers are well-advised to remain cautious, to hold their fire,

81 B. KOCH et al., 13. Journal of European Tort Law 2022, p 44; EUROPEAN LAW INSTITUTE, ‘Guiding Principles for Updating the Product Liability Directive for the Digital Age’, ELI Innovation Paper Series (21 Jan. 2021), pp 9–10.

82 See O. DHEU et al., ‘The European Commission’s Approach to Extra-Contractual Liability and AI – A First Analysis and Evaluation of the Two Proposals’ (6 Oct. 2022), pp 25–27.

83 See also R. DE BRUIN, 7. European Journal of Risk Regulation 2016, pp 490 & 495. 84 M. BUITEN et al., ‘EU Liability Rules for the Age of Artificial Intelligence’ (9 Apr. 2021), p 55. 85 M. BUITEN et al., ‘EU Liability Rules for the Age of Artificial Intelligence’ (9 Apr. 2021), pp 55–56

with further references. 86 B. SCHÜTTE et al., ‘Damages Liability for Harm Caused by Artificial Intelligence – EU Law in Flux’,

Legal Studies Research Paper Series, Paper No 69 (Jan. 2021), p 26. 87 R. DE BRUIN, 7. European Journal of Risk Regulation 2016, p 495. 88 European Commission, ‘Third report on the application of Council Directive on the approximation

of laws, regulations and administrative provisions of the Member States concerning liability for defective products (85/374/EEC of 25 Jul. 1985, amended by Directive 1999/34/EC of the

274

and to resist the urge to legislate, i.e., to sharpen the liability system’.89 More impor- tantly, it also shows that one proposal entails many additional choices to bemade, such as the level at which regulatory initiatives can and should be taken (part 2.3.1.), the specific content of the burden of proof (part 2.3.2.) and the existence of procedural (national) alternatives that may remedy some of the identified concerns (part 2.3.3.).

2.3.1. Level of Intervention

14. A first question relates to the level at which potential initiatives regarding the burden of proof in an AI-context should be adopted. The NewTechnologies Formation (NTF) Expert Report acknowledges that ‘courts might be similarly supportive of victims of emerging digital technologies who have a hard time proving that the technology in question was the actual cause of their harm’.90 At the same time, however, such procedural mechanisms are likely to differ from case to case and most certainly between Member States. Eventually, the NTF Expert Report concludes that the existing differences in procedural rules across Member States could result in different outcomes. This could, therefore, warrant, at least partially, harmonising rules on the burden of proof.91 Yet, it remains unclear whether this should be realized through uniform EU rules or whether the implementation and development of such rules should be left to the Member States’ discretion.92 The EC is clearer in its Report on the Safety and Liability regarding AI as it seeks ‘views whether and to what extent it may be needed tomitigate the consequences of complexity by alleviating/reversing the burden of proof required by national liability rules for damage caused by the operation of AI applications, through an appropriate EU initiative’ (own emphasis).93 The AI Liability Directive also introduces a rebuttable presumption of causality at the supra- national level (cf. Art. 4). Such an EU approach has also been advocated by Lohsse, Schulze and Staudenmayer.94 An EU harmonized initiative helps to ensure the same level of protection for all users in Europe as well as to safeguard a level playing field for

European Parliament and of the Council of 10 May 1999)’, Brussels 14 Sep. 2006, COM(2006) 496 final, p 9.

89 G. WAGNER, Liability for Artificial Intelligence and the Internet of Things, p 46. 90 EXPERT REPORT 2019, p 29–30. 91 EXPERT REPORT 2019, p 30. 92 A. BERTOLINI & F. EPISCOPO, ‘The Expert Group’s Report on Liability for Artificial Intelligence and

Other Emerging Digital Technologies: a critical assessment’, 12. European Journal of Risk Regulation 2021, p 653.

93 European Commission, ‘Report on the safety and liability implications of Artificial Intelligence, the Internet of Things and robotics’, 19 Feb. 2020, COM(2020) 64 final, p 14.

94 S. LOHSSE et al., ‘Liability for Artificial intelligence’, in S. LOHSSE et al. (eds), Liability for Artificial Intelligence and the Internet of Things, p 15; B. SCHÜTTE et al., ‘Damages Liability for Harm Caused by Artificial Intelligence – EU Law in Flux’ (Jan. 2021), p 26.

275

the involved actors, such as producers, providers or operators.95 The idea of a supra- national approach is further strengthened by the fact that there are several other examples in EU legislation introducing a reversal of the burden of proof, such as the Digital Content Directive (DCD) or the Air Passenger Rights Regulation.96

15. At the same time, Buiten notes that the ‘diversity of rules between Member States allows for experimentation and learning, which may be particularly benefi- cial at the beginning of the deployment of a new category of technologies’,97 such as AI. She adds that choosing not to harmonize the rules at the EU level also preserves the coherence of the national tort liability systems.98 Although the adoption of a Regulation may result in less divergence between Member States from the outset, they can (still) be incorrectly interpreted by national courts.99

Buiten and others eventually conclude that an harmonized regime on the European level for AI in general remains ‘questionable’. A form of limited harmonization may on the other hand be justified, especially in those sectors where EU harmonized safety rules already exist and for which supranational harmonized liability rules may a useful complement.100 The targeted approach in the AI Liability Directive cover- ing only one specific aspect of a liability claim – namely causality – clearly shows the reluctance to develop an entire new liability regime encompassing all aspects of tort law.101 In any case, developing a uniform EU approach could be challenging as national law will often remain important in several ways, and hence lead to differences across Member States. One can take the example of the fault-based liability regime in the proposal included in the EP Resolution of October 2020. It relies on concepts such as a caused, reasonable and necessary measures, due diligence or force majeure. The content of these concepts will still have to be applied according to national law, and depending upon the judge’s interpretation, result in

95 This view has been reported and discussed in M. BUITEN et al., ‘EU Liability Rules for the Age of Artificial Intelligence’ (9 Apr. 2021), p 59 or see B. SCHÜTTE et al., ‘Damages Liability for Harm Caused by Artificial Intelligence – EU Law in Flux’, Legal Studies Research Paper Series, Paper No 69 (Jan. 2021), p 26.

96 See for instance Art. 5 of the Air Passenger Rights Regulation (Regulation (EC) No. 261/2004 of the European Parliament and of the Council of 11 Feb. 2004 establishing common rules on compensation and assistance to passengers in the event of denied boarding and of cancellation or long delay of flights, and repealing Regulation (EEC) No. 295/91, OJ L 46; Art. 12 Directive (EU) 2019/770 of the European Parliament and of the Council of 20 May 2019 on certain aspects concerning contracts for the supply of digital content and digital services, OJ L 136.

97 M. BUITEN et al., ‘EU Liability Rules for the Age of Artificial Intelligence’, 9 Apr. 2021, p 59. 98 M. BUITEN et al., ‘EU Liability Rules for the Age of Artificial Intelligence’, 9 Apr. 2021, p 59. 99 B. SCHÜTTE et al., ‘Damages Liability for Harm Caused by Artificial Intelligence – EU Law in Flux’

(Jan. 2021), p 27. 100 M. BUITEN et al., ‘EU Liability Rules for the Age of Artificial Intelligence’ (9 Apr. 2021, p 59. 101 O. DHEU et al., ‘The European Commission’s Approach to Extra-Contractual Liability and AI – A

First Analysis and Evaluation of the Two Proposals’, 6 Oct. 2022, p 14.

276

potentially different approaches.102 More importantly, the AI Liability Directive also relies on concepts such as reasonably likely or duty of care that may need additional interpretation and clarification. Even causality, which is at the core of the Proposal, is a legal term still shaped differently in Member States and not defined in the AI Liability Directive.103 EU instruments sometimes even explicitly leave room for Member States as well. The AI Liability Directive, for instance, does not affect national rules determining which party has the burden of proof, which degree of certainty is required as regards the standard of proof, or how fault is defined, other than in respect of what is provided for in its Articles 3 and 4.104

16. Likewise, Bertolini concludes that national law may still play an important role under the NTF’s suggestions. Although he welcomes the reversal of the burden of proof for defects under the PLD as suggested by the Expert Report from 2019, he also notes that it depends upon ‘disproportionate difficulties or costs’, which are not narrowly defined. If it would be formulated as such in potential legislation, it would leave margin for uncertainty and discrepancies in its daily application.105

Moreover, this provision already presupposes that the victim was able to demon- strate causation. This implies that it must be shown that it is the functioning of the device that resulted in the harmful event which is by far is the most difficult element to prove considering the increasing complexity and interdependence of multiple factors in the causation of harm. Although the Report also addresses the need to alleviate the burden of proof with respect to causation, the single criteria that are proposed may eventually result in diverse applications by national judges as it largely depends upon their subjective assessment.106 There are also different national causation theories in the EU.107 In addition, the NTF only formulates these general considerations on causation while refraining from promoting ‘any specific measure [as it] would run the risk of interfering with national rules of procedure in particular’.108 As such, the most relevant and complex element to be demonstrated in any litigation involving AI would be left untouched. The NTF

102 Compare supra n. 11. Also see J. DE BRUYNE & O. DHEU, ‘Liability for Damage Caused by Artificial Intelligence – Some Food for Thought and Future Debate’, with further references, forthcoming. See for a similar conclusion B. SCHÜTTE et al., ‘Damages Liability for Harm Caused by Artificial Intelligence – EU Law in Flux’ Jan. 2021, p 26.

103 O. DHEU et al., ‘The European Commission’s Approach to Extra-Contractual Liability and AI – A First Analysis and Evaluation of the Two Proposals’ (6 Oct. 2022), pp 17–18 and 21.

104 Article 1.3. AI Liability Directive. 105 A. BERTOLINI, ‘Artificial Intelligence and Civil Liability’ (Jul. 2020), p 82. See also A. BERTOLINI & F.

EPISCOPO, ‘The Expert Group’s Report on Liability for Artificial Intelligence and Other Emerging Digital Technologies: a critical assessment’, 12. European Journal of Risk Regulation 2021, p 653.

106 EXPERT REPORT 2019, pp 49–52. 107 See e.g., M. INFANTINO & E. ZERVOGIANNI, Causation in European Tort Law (Cambridge: Cambridge

University Press 2017), p 726. 108 EXPERT REPORT 2019, p 51.

277

Expert Report also only suggests Member States and their court system to evolve into a given direction. As Bertolini notes, such an approach fails to improve the position of the victims as it may leave to national legislations and judges to decide on a case, potentially leading to various outcomes across Europe. This would evidently lead to increased fragmentation and could possibly decrease legal certainty.109

2.3.2. Content of the Provision

17. This brings us to the second point, namely the content of a potential provision related to the burden of proof or a rebuttable presumption in an AI context. It has already been mentioned that several EU legal instruments contain a reversal of the burden of proof, such as the DCD or the Air Passenger Rights Regulation. These provisions are clear and directly impose the burden of proof upon defendants. Things seem to be more complex when looking at some of the proposals that have already been made regarding AI. They seem to more indirectly impose a burden of proof as it is the judge who eventually has to decide on it.110 With regard to causation, for example, the NTF Expert Report states that the burden of proving causation could be alleviated in the context of challenges raised by emerging technologies ‘if a balancing of some elements warrants doing so’ (own emphasis).111 Some of these elements are also rather subjective, such as the likelihood that the technology at least contributed to the harm and the likelihood that the harm was caused either by the technology or by some other cause within the same sphere.112 The NTF Expert Report also introduces a reversal of the burden of proof regarding the defect under the PLD if there are disproportionate difficulties or costs pertaining to establishing the relevant level of safety or proving that this level of safety has not been met.113 Likewise, Article 9.4 of the revised PLD stipulates that where a national court judges that the claimant faces excessive difficulties, due to technical or scientific complexity, to prove the defectiveness of the product or the causal link between its defectiveness and the damage (or both), the defectiveness of the product or causal link between its defec- tiveness and the damage (or both) can under certain conditions be presumed. It

109 A. BERTOLINI, ‘Artificial Intelligence and Civil Liability’, Jul. 2020, p 82.; A. BERTOLINI & F. EPISCOPO, ‘The Expert Group’s Report on Liability for Artificial Intelligence and Other Emerging Digital Technologies: a critical assessment’, 12. European Journal of Risk Regulation 2021, p 653.

110 See in this regard also the final section of Art. 8.4 of the New Belgian Civil Code, which allows a judge to reverse the burden of proof ‘in light of exceptional circumstances’ when the application of the ‘normal rules’ would be ‘manifestly unreasonable’ (Memorandum of explanation accompanying the Proposed Act to insert Book 8 ‘Evidence’ in the New Civil Code, Parl.St. Kamer 2018–19, no. 3349/001, p 15).

111 EXPERT REPORT 2019, p 49. 112 EXPERT REPORT 2019, pp 49–50. 113 EXPERT REPORT 2019, p 42.

278

remains up to the judge to determine whether there are disproportionate costs or excessive difficulties. This may potentially result in legal uncertainty.

18. As such, if procedural mechanisms were to be introduced by policymakers or further refined in a legal framework, it should be as specific and clear as possible, leaving no uncertainties under which conditions it may actually apply. This would not only be to the benefit of victims but also create legal certainty for AI developers regarding the necessary record-keeping and logging requirements. In this regard, several provisions of the revised PLD are a step in the right direction. Following Article 9.2, the defectiveness of a product, for instance, can be presumed where any of the following conditions are met: (1) the defendant has failed to comply with an obligation to disclose relevant evidence at its disposal pursuant to Article 8(1); (2) the claimant establishes that the product does not comply with mandatory safety requirements laid down in Union law or national law that are intended to protect against the risk of the damage that has occurred; or (3) the claimant establishes that the damage was caused by an obvious malfunction. This seems a clear situation in which the presumption can be applied. Although problematic due to the impor- tance of national law to interpret some of the used concepts, the reversal of the burden of proof included in EP Resolution of October 2020 can be welcomed as well. This resolution suggest that the ‘operator will not be liable if he/she can prove that the harm or damage was caused without his/her fault, relying on either of the following grounds: (a) the AI system was activated without his or her knowledge while all reasonable and necessary measures to avoid such activation outside of the operator’s control were taken, or (b) due diligence was observed by performing all the following actions, namely selecting a suitable AI system for the right task and skills, putting the AI system duly into operation, monitoring the activities and maintaining the operational reliability by regularly installing all available updates’.114

19. One may also wonder whether sectoral approaches specifically tailored to the (needs of actors in a) sector in which the AI system is used may not be considered either.115 This will not only allow to take into account the position of the

114 Article 8 European Parliament Resolution 2020. Also see supra n. 15. 115 The example of third-party certifiers shows that a tailored approach is indeed possible. Different

regulatory regimes (sometimes even including liability provisions) have been adopted for credit rating agencies, auditors and classification societies at the supra – and national level. Scholars have further developed the idea to work with tailored regimes (e.g., A. HAMDANI, ‘Gatekeeper Liability’, 77. Southern California Law Review 2004, p 98) for third-party certifiers by taking into account several criteria, such as appropriate mechanisms inducing certifiers to issue more reliable and accurate certificates (‘triggering mechanisms’), the costs associated with a specific proposal, factors that reduce the risk of a certifier’s potential unlimited liability and a link with existing practices to increase its adoption (see J. DE BRUYNE, Third-Party Certifiers 2019, pp 261–373).

279

claimants, but also the specificities of the sector.116 The NTF Expert Report concludes that ‘no one-size-fits-all solution can (or should) be offered’. Instead, it is ‘necessary to consider a range of options, with the choice within that range to be determined by various factors’.117 The Feasibility Study issued by the Ad Hoc Committee on Artificial Intelligence (CAHAI) also notes that:

because AI has a myriad of applications, remedies need to be tailored towards those different applications. This should include the obligation to terminate unlawful conduct and guarantees of non-repetition, as well as the obligation to redress the damage caused, and compliance with the general rules about the sharing and reversal of the burden of proof in anti-discrimination legislation.118

As noted by Bertolini, a technology specific approach ‘identifying single classes of applications that need to be separately regulated with independent normative acts’ could be a more preferable option.119 This tailored approach would avoid passing over the fundamental differences that exist between different types of AI applications, which each raise specific issues and problems. However, such an approach requires a ‘greater effort’ from the policymakers.120 In this regard, Diamantis already listed a set of criteria that any model of vicarious liability for algorithmic harms should satisfy, namely (1) identify which third party or parties will be liable; (2) be robust enough to avoid gamesmanship; (3) give efficient incentives to the different involved parties; (4) produce fair outcomes; (5) have low barriers to implementation and (6) promote programming values such as interpretability.121 Other potential criteria could be the equal treatment provided to victims, the fair distribution of liability costs (cf. the allocation of liability to parties in relation to their responsibilities in the harm’s occurrence) or the simplification provided by the legal instrument.

20. A final point that may be considered regarding the content of the legal provision relates to the capacity of the plaintiff claiming recovery. Especially

116 See in this regard A. BERTOLINI, ‘Artificial Intelligence and Civil Liability’ (Jul. 2020), pp 87–88, 96–98; O. DHEU & J. DE BRUYNE, ‘European Parliament Study on AI and Civil Liability – Time for the EU to Step up its Game part 1)’ and ‘European Parliament Study on AI and Civil Liability – Towards a “Risk Management Approach” (part 2)’, CiTiP Blog (Aug. 2020); M. BUITEN et al., ‘EU Liability Rules for the Age of Artificial Intelligence’ (9 Apr. 2021), pp 9 and 61; G. SPINDLER, ‘User Liability and Strict Liability in the Internet of Things and for Robots’, in S. LOHSSE et al. (eds), Liability for Artificial Intelligence and the Internet of Things, p 143.

117 EXPERT REPORT 2019, p 36. 118 AD HOC COMMITTEE ON ARTIFICIAL INTELLIGENCE – CAHAI, ‘Feasibility Study’, 17 Dec. 2020, p 38, no.

113. 119 A. BERTOLINI, ‘Artificial Intelligence and Civil Liability’ (Jul. 2020), p 87. 120 A. BERTOLINI, ‘Artificial Intelligence and Civil Liability’ (Jul. 2020), p 87. 121 M. DIAMANTIS, ‘Vicarious Liability for AI’, U Iowa Legal Studies Research Paper No. 2021-27, pp

3–12.

280

consumers – as weaker parties – may/should benefit from the reversal of the burden of proof.122 The idea to work with a reversal of the burden has for instance been discussed under the PLD, which in turn states that ‘protection of the consumer requires that all producers involved in the production process should be made liable’.123 The focus on the position of consumer is of course not surprising in light of other provisions introducing a reversal of the burden of proof to the benefit of consumers, such as the one included in the DCD. Therefore, future legislation may be specifically targeted at consumers, making it more in line with other already existing supranational initiatives.124 Such an approach is especially warranted as the AI Act applies to the ‘user’ of an AI system among others, defining it as ‘any natural or legal person, public authority, agency or other body using an AI system under its authority, except where the AI system is used in the course of a personal non-professional activity’.125 Moreover, even for users, a major shortcoming in the AI Act relates to the lack of rights of redress for individuals as well as to the lack of complaints mechanism. Although the AI Act Proposal is intended to protect funda- mental rights, there is no reference in the text to individuals that are affected by the AI system.126 The AI Liability Directive and the revised PLD include procedural mechanisms to the benefit of claimants, thus providing some effective remedies that can complement the AI Act.127 Moreover, the fact that the AI Liability Directive contains provisions on low-risk AI systems is a step in the right direction as such systems can cause damage as well.

122 By once again referring to the certification sector, legislation as well as case law shows that the (professional) capacity of third parties relying on a certificate is in indeed taken into account as well when determining the liability of third-party certifiers (see e.g., with further references J. DE

BRUYNE, Third-Party Certifiers 2019, p 464). 123 Recital 4 PLD. 124 The Special Committee on Artificial Intelligence in a Digital Age also paves the way for such an approach

when concluding in its report that ‘environments in which AI systems operate may differ in a business-to business (B2B) environment compared to a business-to-consumer (B2C) environment’ (EUROPEAN

PARLIAMENT, SPECIALCOMMITTEE ONARTIFICIAL INTELLIGENCE IN ADIGITAL AGE, ‘Report on artificial intelligence in a digital age’ (5 Apr. 2022), p 35, no. 136). Going beyond the European Union, the Ad hoc Committee on Artificial Intelligence (CAHAI) also addresses this issue in its feasibility report and concludes that ‘[l] iability among business agents could, for instance, be more suitable to address through contractual stipulations rather than through the adoption of a specific liability regime’ (AD HOC COMMITTEE ON

ARTIFICIAL INTELLIGENCE – CAHAI, ‘Feasibility Study’ (17 Dec. 2020), p 45, no. 128). 125 Article 3 (4) Proposal for a Regulation of the European Parliament and of the Council laying down

harmonized rules on artificial intelligence. 126 See e.g., N. SMUHA et al., ‘How the EU Can Achieve Legally Trustworthy AI: A Response to the

European Commission’s Proposal for an Artificial Intelligence Act’ (31 Aug. 2021), pp 44–46, https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3899991.

127 O. DHEU et al., ‘The European Commission’s Approach to Extra-Contractual Liability and AI – A First Analysis and Evaluation of the Two Proposals’ (6 Oct. 2022), pp 15–16.

281

2.3.3. Existing Alternatives

21. Although a reversal of the burden of proof may thus bring benefits in an AI- context, several concerns have already been raised. From a more general per- spective, Koch argues that a reversal of the burden of proof for a defect under the PLD would effectively ‘be nothing less than reallocating the overall risk in disguise’.128 Wagner concludes that is difficult to know whether concerns regarding the reversal of the burden of proof are justified as very few autono- mous products are operating in the market.129 Against this background, one could rely on other procedural mechanisms than a reversal of the burden of proof or working with presumptions to remedy the position of victims.130

22. An example are logging obligations. The AI Act, for instance, requires that high- risk AI systems need to be designed and developed with capabilities enabling the automatic recording of events while the high-risk AI systems is operating. In particu- lar, logging capabilities have to enable themonitoring of the operation of the high-risk AI system with respect to the occurrence of situations that may result in the AI system presenting a risk to the health or safety or to the protection of fundamental rights of person or lead to a substantial modification.131 Buiten rightly suggests that, as AI systems may be equipped with logging or recording mechanisms, the victim should be able to have an easier access to data about the cause of the damage than was previously the case.132 As such, AI systems that were involved in an accident will ‘offer victims, courts and regulators the same comprehensive sets of data that are now available in the case of an airplane’ accident.133 This will, according to some, reduce the evidentiary burden for victims as well as courts.134 In this regard, the disclosure obligations for providers and users in the AI Liability Directive (Art. 3) as well as for defendants in the revised PLD (Art. 8) may be useful.

Bertolini, however, concludes that the most problematic aspect of logged data (and by extension disclosed evidence135) is that its interpretation and analysis might be extremely complicated and costly, especially for the victim. As such, the fact that data is logged and eventually available for the claimant does not appear to be sufficient. To the contrary, he adds that this could lead to improving and

128 B. KOCH, ‘Product Liability 2.0 – Mere Update or New Version?’, in S. LOHSSE et al. (eds), Liability for Artificial Intelligence and the Internet of Things, pp 109–110.

129 G. WAGNER, Liability for Artificial Intelligence and the Internet of Things, pp 45–46. 130 Also see B. KOCH et al., 13. Journal of European Tort Law 2022, p 57. 131 Article 12 Proposal of a Regulation on AI. Also see recommendation [20] in the NTF Expert Report. 132 M. BUITEN et al., ‘EU Liability Rules for the Age of Artificial Intelligence’ (9 Apr. 2021), p 56. 133 G. WAGNER, Liability for Artificial Intelligence and the Internet of Things, p 46. 134 G. WAGNER, Liability for Artificial Intelligence and the Internet of Things, p 46; B. KOCH et al., 13.

Journal of European Tort Law 2022, p 53. 135 O. DHEU et al., ‘The European Commission’s Approach to Extra-Contractual Liability and AI – A

First Analysis and Evaluation of the Two Proposals’ (6 Oct. 2022), pp 20–21.

282

simplifying the producer or operator’s position. Indeed, such actors would have access to a huge amount of data and information which could leave the victim with the uncertain and difficult task of ‘identifying the relevant portion (if it exists) and providing an interpretation of it’.136 Some other changes to regulatory frameworks may also be viable options to remedy the position of victims. A (limited) discovery procedure may, for instance, be introduced when AI systems are involved in in damage inflicting incidents, similar to field-specific ‘discovery rules’ in Articles 5–8 of the Directive dealing with actions for damages under national law for infringe- ments of the competition law provisions of the Member States and of the EU.137

23. National remedies have also been developed to alleviate the burden of proving causation if the claimant’s position is deemed weaker than in typical cases. These can be relied upon in an AI-context as well. Giesen, for instance, discusses the concept of ‘aanvullende stelplicht’ or ‘gemotiveerde betwisting’, which he translates as ‘the duty to provide an extra motivated pleading’.138 A defendant will not only have to deny the plaintiff’s statement of claim and the facts asserted therein but also take an additional step when denying the asserted facts by supplying a certain degree of extra information. This information is typically not available to the plaintiff. By using the ‘aanvullende stelplicht’, the substantiation of a claim is put partly upon the defendant. However, the (legal) burden of proof is not reversed. Only the (evidential) burden of producing evidence is shifted. If the defendant complies with this duty, the plaintiff is still obliged to prove his claim by using the extra provided information.139 The Expert Group Report also refers to procedural options such as the theory of res ipsa loquitur, prima facie evidence and/or low- ering the standard of proof in certain situations or cases.140 Especially lowering the standard of proof has already attracted some attention in an AI-context.141 Belgian civil evidence law provides for a possibility to do so. Article 8.6 of the Belgian Civil Code allows the proof by probability of negative facts and positive facts for which it is impossible or unreasonable to expect proof by the aforementioned reasonable

136 A. BERTOLINI, ‘Artificial Intelligence and Civil Liability’ (Jul. 2020), p 84. 137 Directive 2014/104/EU of the European Parliament and of the Council of 26 Nov. 2014 on certain

rules governing actions for damages under national law for infringements of the competition law provisions of the Member States and of the European Union, OJ L 349 as reported in

B. SCHÜTTE et al., ‘Damages Liability for Harm Caused by Artificial Intelligence – EU Law in Flux’, Legal Studies Research Paper Series, Paper No 69 (Jan. 2021), p 23; M. BUITEN et al., ‘EU Liability Rules for the Age of Artificial Intelligence’ (9 Apr. 2021), p 56.

138 I. GIESEN, European Tort Law 2008, p 59. 139 I. GIESEN, European Tort Law 2008, pp 59–60 with further references. 140 EXPERT REPORT 2019, p 50. 141 See e.g., M. BUITEN et al., ‘EU Liability Rules for the Age of Artificial Intelligence’ (9 Apr. 2021), p

56.

283

certainty standard.142 It is conceivable that cases of tort liability for damage caused by AI systems will be likely candidates for the explicit or implicit use of this probability standard.143 Lowering the burden of proof is also in line with European Court of Justice’s case law regarding the burden of proof under the PLD favouring a more flexible approach to the burden of proof.144

3. Strict Liability in an AI-Context

24. Another example that can be used to illustrate the multi-layered reality of tort law and AI relates to strict liability regimes. Many authors and policymakers have analysed whether or not adopting a strict(er) liability regime for applications embedding AI may be relevant.145 AI systems’ intrinsic characteristics make it challenging for victims to claim compensation, either on the basis of the current PLD or on the basis of national fault-based liability provisions. In a strict liability regime, the victim is no longer required to prove fault, which of course facilitates his/her claim for redress.146 The debate over introducing strict liability rules for AI applications also highlights the question of the purpose that society ascribes to liability mechanisms. In this regard, Zech argues that liability is a risk control tool, and that strict liability should be preferred for novel risks. By relying on law and economics concepts, he posits that strict liability ‘influences the level of care but also the activity level by fully internalising economic risks of AI’.147 He further adds

142 The legislative proposal describes the required probability as 75%, which means that there are serious elements in the file which support the allegations and that alternatives, although not entirely impossible, do not seem credible (Memorandum of explanation accompanying the Proposed Act to insert Book 8 ‘Evidence’ in the New Civil Code, Parl.St. Kamer 2018–19, no. 3349/001, p 17, own translation).

143 J. DE BRUYNE et al., Artificial Intelligence and the Law, p 366. 144 C. DE MEEUS, 4. Journal of European Consumer and Market Law 2019,p 153. See e.g., Cases C-503/

13 and C-504/13 Boston Scientific v. AOK Sachsen-Anhalt and Betriebskrankenkasse [2015] ECLI: EU:C:2015:148; Case C-621/15 W and Others v. Sanofi Pasteur and Others [2017] ECLI:EU: C2017:484.

145 See e.g., C. WENDEHORST, ‘Strict Liability for AI and Other Emerging Technologies’, 11. Journal of European Tort Law 2020, p 179; G. SPINDLER, ‘User Liability and Strict Liability in the Internet of Things and for Robots’ (2019), p 143. On applying strict liability to autonomous vehicle accidents, see A. BERTOLINI & M. RICCABONI, ‘Grounding the Case for a European Approach to the Regulation of Automated Driving: The Technology-Selection Effect of Liability Rules’, 51. European Journal of Law and Economics 2021, p 281; A. ROSENBERG, ‘Strict Liability: Imagining a Legal Framework for Autonomous Vehicles’, 20. Tulane Journal of Technology and Intellectual Property 2017, p 205; Art. 4 European Parliament Resolution 2020. See on the strict liability of developers and backend operators S. LI et al., ‘Liability rules for AI related Harm: Law and Economics Lessons for a European Approach’, European Journal of Risk Regulation 2022, p 16. Also see supra n. 2.

146 See extensively C. VAN DAM, European Tort Law (Oxford: Oxford University Press 2007), pp 255–265.

147 H. ZECH, ‘Liability for AI: Public Policy Considerations’, ERA Forum 2021 (22), p 152.

284

that ‘it also incentivises the further development of existing technologies and, arguably, helps public acceptance. Strict liability may also be used as an instrument for risk-distribution, especially when combined with compulsory liability insurance (third party insurance)’.148

25. Even though strict liability regimes come in many shapes or forms and have been suggested by several scholars and policymakers in an AI-context, such regimes also carry several challenges. For instance, and as opposed to no-fault compensa- tion schemes,149 the injured party may still be required to prove causation (e.g., under the PLD), which can be challenging for highly connected and sophisticated AI systems. Moreover, the qualification of software as a (tangible) product or service falling within the PLD is/was also subject to fierce debates.150 However, the recent EC proposal on a revised PLD does provide some answers to these questions, such as including software within its scope (cf. Art. 4 (1)) or by creating legal presumptions of defectiveness (Art. 9.2). Another identified shortcoming concerning strict liability resides in the so-called ‘chilling effect’ of regulation according to which the prospect of being held strictly liable (e.g., software produ- cers) could slow down or even freeze technological development and innovation.151

Moreover, the specific modalities of a regime may differ illustrating that the choice for strict liability will entail many additional policy choices. Questions, for instance, remain regarding the desired level of intervention (part 3.1.) as well as which parties should eventually be held liable (part 3.2.).152

148 H. ZECH, ‘Liability for AI: Public Policy Considerations’, ERA Forum 2021 (22), p 152. 149 See e.g., K. WATTS, ‘Potential of no-fault comprehensive compensation funds to deal with automa-

tion and other 21st century transport developments’, 12. European Journal of Commercial Contract Law, 2020, pp 1–21; M. SCHELLEKENS, ‘No-fault compensation schemes for self-driving Vehicles’, 10. Law, Innovation and Technology 2018, pp 314–333.

150 See e.g., A. BERTOLINI, ‘Artificial Intelligence and Civil Liability’ (Jul. 2020), p 57; B. KOCH et al., 13. Journal of European Tort Law 2022, p 34–36; G. WAGNER, Liability for Artificial Intelligence and the Internet of Things, p 34; EUROPEAN LAW INSTITUTE, ‘Guiding Principles for Updating the Product Liability Directive for the Digital Age’ (21 Jan. 2021), p 5;C. WENDEHORST, ‘Safety and Liability Related Aspects of Software’, European Commission (Jul. 2021), pp 64–66; Y. BENHAMOU

& J. FERLAND, ‘Artificial Intelligence & Damages: Assessing Liability and Calculating the Damages’ (Mar. 2020), p 9, https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3535387. See however, Art. 4.1 revised PLD.

151 See e.g., S. VAN UYTSEL, ‘Different Liability Regimes for Autonomous Vehicles: One Preferable Above the Other?’, in S. VAN UYTSEL & D. VASCONCELLOS VARGAS, Autonomous Vehicles. Business, Technology and Law (Singapore: Springer 2021), pp (67–92) at p 84 (‘Knowing that there will be always a liability imposed on the manufacturer, innovation could chill. The financial burden may be just too heavy to take the risks of innovation’).

See extensively M. SCHELLEKENS, ‘Self-driving cars and the chilling effect of liability law’, Computer Law & Security Review 2015, pp 506–517.

152 The following article analyses some of these interrogations, e.g., M. BUITEN et al., ‘EU Liability Rules for the Age of Artificial Intelligence’ (9 Apr. 2021). In reviewing the possible EU liability

285

3.1. Level of Intervention

26. Similar to the procedural elements such as the reversal of the burden of proof, policymakers will need to determine the scope and level at which legal intervention should occur. This entails two aspects. On the one hand, it requires to establish whether a strict liability regime should be considered at a national or an EU level (part 3.1.1.). On the other hand, it also implies determining whether the specific regime should be considered on a horizontal level (and/or be technologically neutral) or be framed on a sector-by-sector approach (part 3.1.2.).

3.1.1. National Versus European Approach

27. As has already been mentioned, both an EU and national approach have their benefits. A national-based approach allows to not entirely overhaul the liability frameworks already in place. It would, therefore, prevent complexification and also potentially avert any legal incoherence’s and unbalances from occurring due to the uncertain interaction between a European legal instrument and existing national strict liability regimes.153 Indeed, one of the issues in adopting a European regime resides in the fact that it may still need to rely on national based frameworks (e.g., for the interpretation of concepts such as ‘reasonableness’, ‘force majeure’ or ‘reasonably likely’).154 This could lead to situations in which legal certainty could possibly be jeopardized. Moreover, a national approach allows jurisdictions to rely on the existing body of law, including case law, and adapt it to the local context. It can also be noted that some jurisdictions already have strict national-based liability regimes that could potentially be applicable to some AI-applications, such as autonomous vehicles.155 This, for example, is (at least partially) the case for the strict road traffic liability mechanisms in Germany and France.156 In the event of

frameworks, it looks at ‘(1) who should be liable; (2) the scope of new rules; and (3) the level of harmonization’ (p 8).

153 See e.g., J. DE BRUYNE & O. DHEU, ‘What are the EU’s orientations and envisaged choices for the regulation of liability and artificial intelligence?’, KCDS Blog (22 May 2020). Also see supra nn. 14–16.

154 J. DE BRUYNE & O. DHEU, ‘Liability for Damage Caused by Artificial Intelligence – Some Food for Thought and Future Debate’ (2022), forthcoming.

155 See on the discussion whether existing road liability law regimes are adequate for dealing with damages arising out of autonomous vehicles and on the unequal treatment of current road traffic liability laws E.F.D. ENGHELHARD & R. DE BRUIN, ‘EU Common Approach on the liability rules and insurance related to Connected and Autonomous Vehicles’, Annex I, European Parliament (Feb. 2018), pp 69–84.

156 See for Germany e.g., F. PÜTZ et al., ‘Reasonable, Adequate and Efficient Allocation of Liability Costs for Automated Vehicles: A Case Study of the German Liability and Insurance Framework’, 9. European Journal of Risk Regulation 2018, p 548; M. EBERS, ‘Civil Liability for Autonomous Vehicles in Germany’ (5 Feb. 2022) (https://ssrn.com/abstract=4027594). See for France e.g., I. VINGIANO, ‘Quel avenir juridique pour le “conducteur” d’une “voiture intelligente” ?’, Lextenso,

286

damages caused by an autonomous vehicle, some jurisdictions do not require the victims to identify the potential wrongdoer and prove fault.157 They can usually continue to bring a claim against the vehicle keeper or user whose vehicle is insured for third party liability. Allocating final liability and associated costs will usually be left to the vehicle insurer through recourse action.158

28. At the same time, however, a national approach also has shortcomings. Indeed, it could lead to legal fragmentation between Member States laws, which could be detrimental to the legal safety of users of AI driven products or services.159

This is particularly true for transport mediums as the possibility of cross border accidents and the lack of harmonized approach could push in favour of adopting a

Petites Affiches 2014, p 6. Borghetti argues that strict road traffic liability regimes which are based on the use (cf. Germany) or the ‘happening of an accident’ (cf. France) could still apply (J.-S. BORGHETTI, ‘Civil Liability for Artificial Intelligence: What Should Its Basis Be?’, 17. Revue des juristes de Sciences Po 2019, pp 94–102).

157 For instance, in France, the 1985 Badinter law on the compensation of road traffic victims only requires the injured party to prove that he/she suffered a damage in relation to a road traffic accident in which a terrestrial motor vehicle was involved. The victim is not required to prove the driver’s fault. (Art. 1 Loi n° 85–677 du 5 juillet 1985 tendant à l’amélioration de la situation des victimes d’accidents de la circulation et à l’accélération des procedures d’indemnisation). In Germany, the general road traffic rules could apply to autonomous vehicle accidents (see the Road Traffic Act or Straßenverkehrsgesetz (StVG)). An amendment to this Road Traffic Act was adopted in 2017, adding special provisions regarding autonomous vehicles, but without over- hauling the general liability framework. In 2021 the Road Traffic Act was further amended. The keeper of the vehicle is held strictly liable, but the driver can also be held liable for presumed fault. The UK also adopted the Automated and Electric Vehicle’s Act 2018 (‘AEVA’), which makes the vehicle insurer strictly ‘liable’ for damages arising out of the use of vehicles in automated mode.

158 This depends on the applicable road traffic regime. If it is not based on fault, the victim can usually bring a claim against the vehicle insurer who may in return have recourse rights against liable third parties. See E.F.D. ENGHELHARD & R. DE BRUIN, ‘EU Common Approach on the liability rules and insurance related to Connected and Autonomous Vehicles’, Annex I, European Parliament (Feb. 2018), pp 95, 111–115.

159 In this regard, a comparison can once again be made with the legal situation of third-party certifiers. Liability regimes have been adopted at the EU level covering the activities of credit rating agencies or classification societies acting on behalf of flag states (so-called Recognized Organizations). Some of the reasons that are mentioned as to why an EU regime is appropriate for third-party certifiers relate to (1) their societal importance (2) the lack of contractual relation- ship between third parties and certifiers making it difficult to establish liability, and (3) the influence of the divergence between Member States regarding liability on the implementation of applicable EU legislation. Arguably, similar arguments can be made in the context of AI as it also has an impact on the wider society (see e.g., the AI Liability Directive stipulating that ‘In terms of social impacts, the Directive will increase societal trust in AI-technologies and access to an effective justice system’ (p 4, also see Recital (4)).

287

European legal instrument.160 Moreover, Wendehorst argues that victims should be treated equally if they are harmed by the same dangers.161 This would not be achieved through a national-based approach in which the outcome of the victim’s claim would greatly differ between jurisdictions. European legal instruments would assure harmonization of the law across Europe and allow victims to be equally treated throughout jurisdictions. The legal basis for EU intervention could be Article 114 of the Treaty of the Function of the EU (TFEU) and could take the form of either a Regulation or a Directive.162

3.1.2. Horizontal and/or Vertical Regimes

29. Another question relates to whether strict liability in an AI-context should be horizontal or, instead, vertical, which entails a more sectoral approach. Bertolini concludes that the EU should continue to adopt a sectoral and technology specific approach to regulation as he posits that a uniform regulation for all AI based applications is not needed, even less for liability.163 AI is and will be used in many fields including ‘transportation, medical diagnosis, capital markets, consu- mer products and services, industrial production, energy production and distribution’.164 Considering that even liability aspects are largely ‘separately regu- lated’, ‘they should continue to be separately regulated when AI-based solutions are implemented’.165 Such a technology-specific and application-specific approach to the regulation of AI also ‘better conforms to the principles of proportionality and subsidiarity, minimizing risks of undesirable interferences with Member States’ legal systems’.166 Moreover, AI technologies are used as tools in many tangible products and are not always standalone items.

30. By contrast, the Resolution issued by the EP in October 2020 contains a horizontal approach for AI systems.167 It imposes a strict liability regime for opera- tors of high-risk AI systems. High-risk refers to a significant potential in an

160 E.F.D. ENGHELHARD & R. DE BRUIN, ‘EU Common Approach on the liability rules and insurance related to Connected and Autonomous Vehicles’, Annex I, European Parliament (Feb. 2018), pp 95–96.

161 C. WENDEHORST, ‘Strict Liability for AI and Other Emerging Technologies’ (2020), p 173. 162 See for a nuanced view on Art. 114 and an analysis on full or maximum harmonization legal

instruments M.B.M. LOOS, ‘Harmonization of Private Law in the 2020s: Targeted Full Harmonization 2.0?’, Amsterdam Law School Legal Studies Research Paper n°2022-14 (2022).

163 See on the need of a technology specific approach: A. BERTOLINI, ‘Artificial Intelligence does not exist: defying the technology-neutrality narrative in the regulation of civil liability for advanced technologies’ (2022), pp 414–416.

164 A. BERTOLINI, ‘Artificial Intelligence and Civil Liability’ (Jul. 2020), p 63, no 40. 165 A. BERTOLINI, ‘Artificial Intelligence and Civil Liability’ (Jul. 2020), p 63, no 40. 166 A. BERTOLINI, ‘Artificial Intelligence and Civil Liability’ (Jul. 2020), p 63, no. 40 and 41. 167 J. DE BRUYNE & O. DHEU, ‘What are the EU’s orientations and envisaged choices for the regulation

of liability and artificial intelligence?’, KCDS Blog (22 May 2020).

288

autonomously operating AI system to cause harm or damage in a way which is random and that goes beyond what can reasonably be expected. The significance of the potential depends on the interplay between the severity of possible damage, the degree of autonomy of decision-making, the likelihood that the risk materializes and the manner and the context in which the AI system is used.168 All high-risk AI systems and all critical sectors in which they are used need to be listed in the Annex to the Regulation.169 The Resolution proposes a fault-based liability regime with a reversal of the burden of proof for operators of AI systems that do not constitute a high-risk AI system and are not listed in the Annex.170 Although not laying down a full-fledged liability regime for AI systems, the proposed AI Liability Directive does tip on the side of a more horizontal approach applying to a variety of sectors as well.171

Whereas such an approach facilitates the victim’s claim for compensation (cf. single point of entry, single set of rules, etc.), the horizontal liability regime in the EP Proposal (and by extension the AI Liability Directive) may not be the perfect match after all as it imposes a ‘one size fit all’ regime. In the case of the EP proposal, it establishes a single set of liability caps,172 which fails to consider the specifics of each sector.173 First and foremost, the risks are not necessarily the same between different classes of AI applications. For instance, one may wonder whether the damage risks of an autonomous unmanned aircraft operating in a dense urban environment are the same as those of an autonomous vehicle, or even that of an AI-driven medical diagnosis software. This question is also related to the liability caps that the proposed EP Regulation establishes for high-risk AI applications.174 The proposed EP Regulation sets the liability caps to two million euros ‘in the event of the death of, or in the event of harm caused to the health or physical integrity of, an affected person, resulting from an operation of a high-risk AI-system’.175 The liability cap is set to a maximum amount of one million euros ‘in the event of significant immaterial harm that results in a verifiable economic loss or of damage caused to property, including when several items of property of an affected person were damaged as a result of a single operation of a single high-risk AI-system’.176

168 Article 3 (c) European Parliament Resolution 2020. 169 Article 4.2 European Parliament Resolution 2020. 170 Articles 8.1 and 8.2 European Parliament Resolution 2020. 171 See O. DHEU et al., ‘The European Commission’s Approach to Extra-Contractual Liability and

AI – A First Analysis and Evaluation of the Two Proposals’ (6 Oct. 2022), p 41. 172 Article 5 European Parliament Resolution 2020. 173 See on the issue of liability caps imposed by the EP proposal A. BERTOLINI, ‘Artificial Intelligence

does not exist: defying the technology-neutrality narrative in the regulation of civil liability for advanced technologies’ (2022), pp 407–412.

174 Article 5 European Parliament Resolution 2020. 175 Article 5.1 (a) European Parliament Resolution 2020. 176 Article 5.1 (b) European Parliament Resolution 2020.

289

Although this may be adequate for certain classes of AI applications (e.g., autono- mous vehicles), it may not necessarily be the case for other classes of AI applications (e.g., unmanned passenger aircraft flying in a dense urban environment or an AI- driven medical diagnosis software). Another issue in the EP proposed Regulation is that it risks making the injured party better compensated under national fault-based liability law (for non-high-risk AI applications where liability is unlimited) rather than under the strict liability regime for high-risk AI applications (where there are liability caps, which may seem artificially low). This seems paradoxical.177

3.2. Liable Party/Parties

31. Another fundamental question resides in determining against whom strict liability should be channelled and whether it should be focused on the user, the operator, the manufacturer or any other involved party. The answer is highly dependent on the social values, the choice of normative criteria and the societal priorities that the legislator decides upon.178 It may also depend on the sector in which it applies if a sectoral approach would be considered.179 In any case, clear normative criteria need to be determined in order to justify channelling liability against one or another party. Such criteria could include ‘legal certainty’ provided by the legal regime, ‘simplicity’, ‘easiness in application’, ‘adequate compensation’, etc.180 Moreover, a clear scoping exercise will have to be carried out when defining the potential liable party. Adopting a narrow or broad definition will have differ- entiated results/impacts on the position of these actors. Taking this into account, we will first focus on the manufacturer regardless of how that actor is eventually defined/determined (part 3.2.1.). The operator is very often considered to be a potential party as well that can be held strictly liable (part 3.2.2.) However, we will explore the limitations of such an approach and look at possible alternatives (part 3.2.3.).

177 D. GALBOIS-LEHALLE, ‘Responsabilité Civile Pour l’intelligence Artificielle Selon Bruxelles: Une Initiative à Saluer, Des Dispositions à Améliorer’, 2. Recueil Dalloz 2021, pp 87–88; A. BERTOLINI, ‘Artificial Intelligence does not exist: defying the technology-neutrality narrative in the regulation of civil liability for advanced technologies’ (2022), p 409.

178 C. SIYUAN, ‘Regulating autonomous vehicles: Liability paradigms and value choices’, in AI, Data and Private Law 2021, Research Collection School Of Law pp 147–172, https://ink.library.smu. edu.sg/sol_research/3403.

179 A. BERTOLINI, ‘Artificial Intelligence and Civil Liability’ (Jul. 2020), pp 31 and 61. See also quoted author M.WHITTAKER et al., ‘AI Now Report 2018’ (2018, AI Now Institute, New York University), p 8.

180 Also see C. WENDEHORST, ‘Safety and Liability Related Aspects of Software’ (Jul. 2021), p 96.

290

3.2.1. The Manufacturer?

32. A common assumption resides in the prediction that AI-driven systems will eventually lead to a transfer in control from the user to the machine,181 which could therefore lead to a liability shift from the user to the manufacturer.182 This is based on the premise that AI-driven systems will become increasingly autonomous meaning that human intervention and inputs will lose importance.183 This may push the focus on making the manufacturer liable for damages caused by the AI system or device based on product liability.

33. However, existing product liability laws, both at the European and national level, may be difficult to apply in the context of AI-driven systems.184 Referring to autonomous (transportation) mediums, Dheu, Ducuing and Valcke argue that a so- called ‘product-oriented paradigm’, defined ‘as the regulatory focus placed on product legislation as the main tool leveraged to regulate the “New Vehicle”’ would not necessarily fit within the foreseen technological and operational eco- system surrounding such AI driven systems.185 Autonomous vehicles, and AI-driven systems in general, are blurring the lines between products and services.186 Indeed, the vehicle itself relies on frequent software updates and upgrades that will require constant monitoring from the manufacturer (and other related parties). As such, the manufacturer’s responsibilities no longer limit themselves to designing and producing a product but could also extend to the vehicle’s dynamic lifecycle. The manufacturing and operational phases are, as noted by some, ‘getting intertwined in time’.187 This could result in a possible risk of loss of coherence and blurriness in distinguishing between the manufacturer’s liability based on the PLD and the

181 G. WAGNER, Liability for Artificial Intelligence and the Internet of Things, pp 37–38. He develops the notion of control shift from the user to the manufacturer within the more general context of robotics.

182 See on the projected liability shift in the context of autonomous vehicles N. KALRA et al., ‘Liability and Regulation of Autonomous Vehicle Technologies. California PATH Research Report’ (RAND corporation 2009), pp 14 and 22; M.F. LOHMANN. ‘Liability Issues Concerning Self-Driving Vehicles’, 7. European Journal of Risk Regulation 2016, pp 337–338.

183 On the control shift, see for instance A. GALASSO & H. LUO, ‘Punishing Robots: issues in the economics of tort liability and innovation in artificial intelligence’, in A. AGRAWAL et al. (eds), The Economics of Artificial Intelligence: An Agenda (University of Chicago Press 2018), p 493.

184 See for a discussion and further references J. DE BRUYNE et al., in Artificial Intelligence and the Law, pp 376–388; W. BARFIELD, ‘Liability for autonomous and artificially intelligent robots’, 9. Paladyn, Journal of Behavioral Robotics 2018, pp 196–197.

185 O. DHEU et al., ‘The Emperor’s New Clothes: A Roadmap for Conceptualizing the “New Vehicle”’, 130. Revue TRANSIDIT (2020), p 14.

186 M. BUITEN et al., ‘EU liability rules for the age of Artificial Intelligence’ (2021), p 51. 187 O. DHEU et al., 130. Revue TRANSIDIT (2020), p 14; See quoted reference: EXPERT REPORT 2019,

pp 39 and 44, where they seem to evoke the possibility of the manufacturer acting as an ‘operator’. They differentiate between the ‘frontend’ operator from the ‘backend’ operator.

291

manufacturer’s liability for over the air (‘OTA’) software updates possibly based on fault liability. This could create increased complexity and confusion for victims and liable parties. In the context of fully autonomous vehicles, it could make sense to channel liability against the manufacturer, but on a new legal basis. The revised PLD ‘overstretches’ the product liability regime and extends it to actors and items (understood in the broader term) not related to products stricto sensu. This can be seen both in the clear inclusion of software within its scope as well as in the reference to ‘related services’ and ‘any item’ in the definition of a component. This de facto inclusion of certain product-related services, whether provided by the manufacturer or a third party, has the effect of extending the scope of the product liability regime and associated (legal) responsibilities. This confirms the ‘product- oriented paradigm’ pattern.188

Though referring to high-risk AI systems in general (not only autonomous vehicles), the EP’s 2020 Resolution leans in that direction. Indeed, it foresees the possibility of applying a specific strict liability regime to operators of so-called high- risk AI systems (and therefore autonomous vehicles) including ‘backend operators’. This could cover manufacturers, in their quality of defining ‘the features of the technology, [providing] data and essential backend support service and therefore also [exercising] a degree of control over the risk connected with the operation’. Note, however, that the Resolution does not directly mention manufacturers as backend operators.189 Yet, the distinction between the manufacturer acting as a producer and the manufacturer acting as an operator is not that clear, as demon- strated by the Resolution’s seemingly conflicting Recital (12) and its annex’s Article 4.4,190 even though Article 11 contains some level of articulation between the PLD and the proposed liability regime for operators (cf. infra no. 36).

3.2.2. The Operator?

34. The manufacturer is not the only party involved in the functioning of the AI system. Article 7 of the revised PLD lists the types of ‘economic operators’ that can be held liable for defective products, which is broader than the previous one. It will apply to manufacturers, component manufacturers of tangible or intangible or any

188 O. DHEU et al., ‘The European Commission’s Approach to Extra-Contractual Liability and AI – A First Analysis and Evaluation of the Two Proposals’ (6 Oct. 2022), pp 37–38.

189 See Arts 3 (f), 4.1 and 4.4 of European Parliament Resolution 2020. 190 Recital (12) of European Parliament Resolution 2020 states that it ‘considers it appropriate to

understand “operator” to cover both the frontend and backend operator, as long as the latter is not covered by the PLD’ (own emphasis). Art. 4.4 proposes that ‘[t]he backend operator shall ensure that its services are covered by business liability or product liability insurance’ (own emphasis). It is, therefore, not clear what the boundaries of product liability versus strict liability of the backend operator are under this proposed regime and whether this regime would include manufacturers in the scope of backend operators.

292

related service that is integrated or interconnected with a product, the importer and authorized representative, fulfilment service providers or online platforms allowing consumers to conclude distance contracts with traders and that are not qualified as a manufacturer, importer or distributor. At the same time, however, no (explicit) reference is made to backend operators even though they seem to serve similar functions as manufacturers. In any case, the notion of ‘economic operator’ is rather confusing as it seems to indicate that the regime extends well beyond the realm of pure manufacturing. This would significantly affect the very nature of the Directive, which is no longer specifically targeted at products and manufacturing per se. Such modifications require a well-thought analysis of the long-tail conse- quences. It is also crucial to maintain alignment within the various branches of product legislation, in the process of adapting it to new technologies.191

Indeed, many AI-driven devices will be operated by private and/or profes- sional/commercial operators. Many reasons plead in favour of holding the operator liable under certain circumstances. It has been noted that most current AI systems are not fully autonomous and still require human participation or control.192 This means that it is still advisable to hold the human operator liable to induce safe behaviour and decision-making.193 Even in fully autonomous systems, the operator (still) continues to decide over its use.194 This would act again as an incentive for the operator to properly maintain the AI system and update the safety critical software. Moreover, since they benefit from the use of such AI driven systems and bear the risk of their operation, they could also be held strictly liable for damages arising out of their use as the single point of entry for the victim. Creating such a legal regime would clearly facilitate the victims’ claim process as the operator would be identifi- able and the plaintiff would not be required to prove any fault.195

35. At the same time, however, several challenges arise.196 Ascribing liability to the operator in all circumstances may not be advisable.197 In this regard, Bertolini

191 O. DHEU et al., ‘The European Commission’s Approach to Extra-Contractual Liability and AI – A First Analysis and Evaluation of the Two Proposals’ (6 Oct. 2022), p 35.

192 M. BUITEN et al., ‘EU liability rules for the age of Artificial Intelligence’ (9 Apr. 2021), p 56. 193 See as quoted by BUITEN et al.: A. GALASSO & H. LUO, ‘Punishing Robots: issues in the economics of

tort liability and innovation in artificial intelligence’, in A. AGRAWAL et al. (eds), The Economics of Artificial Intelligence: An Agenda (University of Chicago Press 2018), pp 493–504.

194 M. BUITEN et al., ‘EU liability rules for the age of Artificial Intelligence’ (9 Apr. 2021), pp 56–57. 195 See on the point of facilitated compensation of victims e.g., EXPERT REPORT 2019, pp 36, 39–42. 196 VAN UYTSEL, for instance, acknowledges that a strict liability regime targeting the vehicle operator

would enable the victim to have a one stop window. However, with connected and cooperative vehicles, the operator’s insurer may want to address recourse action against other liable parties such as the manufacturer, the service providers, infrastructure manager. This could result in costs being shifted to the operator (S. VAN UYTSEL, ‘Different Liability Regimes for Autonomous Vehicles’ (2021), p 89).

197 See e.g., C. WENDEHORST, ‘Strict Liability for AI and Other Emerging Technologies’ (2020), p 178.

293

favours a risk-management based approach, which implies that liability should be attributed to the party best placed to ‘(1) identify a risk (2) control and minimize it through its choices, and (3) manage it – ideally pooling and redistributing it among all other parties – eventually through insurance, and/or no fault compensation funds’.198 Yet, the operator may not always have much control over the system that it operates. For instance, in the case of fully autonomous vehicles (if they would ever be allowed on public roads at all199), at most, the operator may decide over the direction and destination of the vehicle as well as choose to activate it or not. This may challenge the alleged deterrent effect of liability when applied to the operator. Indeed, although this has been contested,200 one of liability’s alleged goals is to incentivize safe behaviour of conducts. The question arises as to what would happen when the potential liable party has little, if any, margin of adapting the system’s behaviour into that of a safer one. As such, the deterrent effect of liability may be undermined.

36. Another problem relates to delineating and defining the operator of an AI system, which may be troublesome. As indicated above, the EP Resolution con- tains a liability regime for operators of AI systems. It does not refer to the notion of ‘user’ but to that of the ‘operator’, which is much broader.201 Two categories of operators are distinguished: the frontend operator and the backend operator.202

Frontend operators primarily decide and benefit from the use of the technology (e.g., a user of autonomous vehicle). Backend operators are those actors who ‘continuously define the features of the relevant technology and provide essential

198 A. BERTOLINI, ‘Artificial Intelligence and Civil Liability’ (Jul. 2020), p 99. 199 See in this regard the many contributions in J. DE BRUYNE (ed.), Autonome motorvoertuigen. Een

multidisciplinair onderzoek naar de maatschappelijke impact (Bruges: Vanden Broele 2021), p 405. 200 Behavioural law and economics scholars, for instance, question the underlying rational choice

assumptions and endeavour to render economic analysis more realistic by using psychological insights (e.g., K. MATHIS, European Perspectives on Behavioural Law and Economics (SpringerLink 2015), p 271; C. JOLLS et al., ‘A Behavioral Approach to Law and Economics’, 50. Stanford Law Review 1998, pp 1471–1550). Several (empirical) studies even show that tort law does not always have the expected deterring influence on someone’s behaviour (e.g., D.W. SHUMAN, ‘Psychology of Deterrence in Tort Law’, 42. University of Kansas Law Review 1993, p 165; J.W. CARDI et al., ‘Does Tort Law Deter Individuals? A Behavioral Science Study’, 9. Journal of Empirical Legal Studies 2012, p 567).

201 See Art. 1 and Art. 3 (d) European Parliament Resolution 2020. See for an extensive analysis of the (too) broad definition of both frontend and backend operators and their articulation with the PLD A. BERTOLINI, ‘Artificial Intelligence does not exist: defying the technology-neutrality narrative in the regulation of civil liability for advanced technologies’ (2022), pp 401–407. This author looks at the multiplicity of actors liable as ‘operators’ as well as ‘complex interactions between them’. See on the heterogeneity of operators and the blurred distinction between ‘operation’ and ‘creation’ S. LI et al., ‘Liability rules for AI-related Harm: Law and Economics Lessons for a European Approach’ (2022), p 11. Also see supra n. 30.

202 See Art. 3 (e) and (f) European Parliament Resolution 2020.

294

and ongoing backend support’ (e.g., the parties that provide continuous software updates as well as backend services).203 In essence, looking at the EP Resolution, the producer’s liability seems to extend outside the realm of the PLD as the backend operator’s definition possibly includes the manufacturer. The ‘camou- flaging function’ of the backend operator204 could serve the purpose of extend- ing strict liability to cases in which the manufacturer continues to provide updates/upgrades to the product. But as Wagner explains, this may duplicate product liability into two separate regimes.205 Though this could be helpful from the perspective of the victim as he/she would have two separate legal basis’ for bringing a claim, it would have the effect of rendering even more complex an already blurry liability setting. Moreover, the manufacturer would see increased liability exposure, which could be detrimental to innovation. Nevertheless, Article 11 of the Resolution does contain certain rules with regard to the articulation between the strict operator centric liability and the product liability regime. Indeed, it states that if the if ‘the backend operator also qualifies as a producer as defined in Article 3 of the PLD, that Directive should apply to him or her’.206

One may also wonder whether a narrow(er) definition of ‘operator’ than the one used in the current EP Resolution should be considered. Indeed, the uncertain interplay between the producer’s strict liability falling under the definition of ‘backend operator’ in the EP Proposal and that of the producer’s liability under the current PLD can be questioned. By focusing on the operator, understood as the party that uses and exploits the system, the victim could more clearly differentiate between the end user’s liability (operator stricto sensu) and that of the manufacturer (under the PLD). Instead of such a broad definition, the operator could be apprehended as the party (natural or legal person) who operates an AI-driven system for its own benefit or for that of others. In the context of AI-driven transportation systems, a further distinction could also be made between professional operators which would be subject to strict liability

203 EXPERT REPORT 2019, pp 39–41. The Report states that when ‘there is more than one operator, such as a frontend and a backend operator’, that ‘strict liability should be on the one who has more control over the risks posed by the operation’ (p 41). It also recommends that the backend operator should be held primarily liable as he/she is ‘in a position to control, reduce and insure the risks associated with the use of the technology’ (pp 41–42). See also O. DHEU, ‘EU report on AI, new technologies and liability: key take-aways and limitations’, CiTiP Blog (9 Jan. 2020).

204 G. WAGNER, ‘Liability for Artificial Intelligence: A Proposal of the European Parliament’ (14 Jul. 2021), p 11 (available on SSRN), published in H. EIDENMÜLLER & G. WAGNER, Law by Algorithm (Tübingen: Mohr Siebeck 2021), p 272.

205 See on this point and for a comparison with the PLD G. WAGNER, Law by Algorithm, pp 11–12. 206 See on this point A. BERTOLINI, ‘Artificial Intelligence does not exist: defying the technology-

neutrality narrative in the regulation of civil liability for advanced technologies’ (2022), p 407. This author is critical of the articulation in the proposal between the PLD and the liability of the operator(s).

295

and private operators not subject to such a stringent liability mechanism.207

This would be justified by the level of information/knowledge that the profes- sional operator has over his or her system. In any situation, the operator (or his/her insurer) should be invested with effective recourse rights against liable third parties in order to recover part or all costs related to the accident. This is important in order to have a fair or effective apportionment of liability costs and also foster safe conducts from all involved parties.

3.2.3. Other Parties?

37. As has been rightfully noted in the Expert Group’s Report, the sheer numbers of potential liable parties could make it challenging for the victim to successfully bring a claim against the proper liable party.208 Moreover the ‘many hands’ issue further complicates the allocation of liability.209 This problem arises because the development and operation of AI systems typically entail contributions from var- ious ‘individuals, organizations, machine components, software algorithms and human users, often in complex and dynamic environments’.210 Joint and several liability could help alleviate some of these difficulties for the victim.211

However, many other parties could potentially also be held strictly liable for the damages suffered by the victim. Such parties range from the software program- mer, the network providers, the data providers to the public authorities, for instance in their capacity of infrastructure managers. The AI Liability Directive also includes a presumption of causality that applies to different parties, such as ‘defendants’ as well as providers of high-risk AI systems, persons acting on the latter’s behalf or users.212 Wagner clusters relevant parties between those respon- sible for the creation of autonomous systems on the one hand (‘manufacturers’) and

207 On not subjecting frontend operators who are consumers to strict liability, see C. WENDEHORST, ‘Strict Liability for AI and Other Emerging Technologies’ (2020), p 178. Also see H. ZECH, ‘Liability for AI: public policy considerations’, ERA Forum 2021 22(1), p 156 ‘A strict liability for operators of high-risk AI systems should only be introduced for operators with special expertise (operators whose main business purpose is the operation of digital systems)’ and A. BERTOLINI, ‘Artificial Intelligence does not exist: defying the technology-neutrality narrative in the regulation of civil liability for advanced technologies’ (2022), p 407 (‘an obvious distinction that is for instance neglected [in the EP proposal] is that between operators acting on personal of professional grounds’).

208 EXPERT REPORT 2019, p 56. 209 M. DASTANI & V. YAZDANPANAH, ‘Responsibility of AI systems’, AI & Society (2022), s. 6. 210 COUNCIL OF EUROPE (Expert Committee on human rights dimensions of automated data processing

and different forms of artificial intelligence (MSI-AUT)), ‘Responsibility and AI’, 2019, DGI(2019) 05, pp 11 and 62–64.

211 See in this regard also Art. 11 European Parliament Resolution 2020. 212 Article 4 AI Liability Directive.

296

those responsible for operating them (‘users’) on the other hand.213 Some authors such as Reed, when referring to automated vehicles, argue in favour of making the keeper strictly liable which ‘follows the precedent for aircraft owners’.214 This comes close to existing liability regimes for custodians of defective things, such as the one in Article 1384 Old Civil Code in Belgium.215 Moreover, it can be noted that the UK has already adopted a specific piece of legislation which applies to (certain) automated vehicles: the 2018 Automated and Electric Vehicles Act. This law makes the vehicle liability insurer strictly liable for damage caused by the vehicle when driving in automated mode.216 In this sense, neither the operator (or user) nor the manufacturer are ‘liable’ to compensate under this law (at least directly). Despite this piece of anticipatory regulation, certain issues have been raised.217 In any case, the liable person or entity may have a recourse action against liable third parties, such as the manufacturer.

4. Concluding Remarks

38. This article relied on two use cases to show the ‘multi-layered’ reality regard- ing tort liability and AI. It was shown that policy decisions in the field of the allocation of the burden of proof and the adoption of strict liability regimes actually entail several additional choices to be made, such as the level of legal intervention (cf. national vs European, horizontal vs sectoral), its content as well as those parties that would be subject to these (new) rules. This does not only show the complexity and the many challenges in the field of tort law and AI but also that there is currently no ‘perfect’ liability mechanism that can help address all the challenges raised by AI systems. The AI Liability Directive and the revised PLD provide some answers to these challenges. However, it remains to be seen how these proposals will evolve as they will be discussed and amended by the EP and the Council.

213 G. WAGNER, Liability for Artificial Intelligence and the Internet of Things, p 32 and further. 214 C. REED et al., ‘Responsibility, Autonomy and Accountability: Legal Liability for Machine

Learning’, Microsoft Cloud Computing Research Centre 3rd Annual Symposium, London (9 Sep. 2016). p 29.

215 See e.g., J. DE BRUYNE & J. TANGHE, ‘Liability for Damage Caused by Autonomous Vehicles: A Belgian Perspective’, 8(3). Journal of European Tort Law 2017, pp 384–355; O. DHEU et al., ‘Autonomous Vehicles and Civil Liability in Belgium’, in B. VON BODUNGEN et al. (eds), Autonomous Vehicles and Civil Liability in a Global Perspective (Heidelberg: Springer 2023), forthcoming.

216 Section 2 of part 1 of the Automated and Electric Vehicles Act (AEVA) 2018. 217 See for an analysis of the AEVA M. CHANNON, ‘Automated and Electric Vehicles Act 2018: An

Evaluation in Light of Proactive Law and Regulatory Disconnect’, 10. European Journal of Law and Technology 2019, p 37; F. PAOLO PATTI, ‘The European Road to Autonomous Vehicles’, 43. Fordham International Law Journal 2019, p 125; J. MARSON et al., ‘The Automated and Electric Vehicles Act 2018 Part 1 and Beyond: A Critical Review’, 41. Statute Law Review 2019, p 359.

297

39. Finally, although this article limited itself to liability mechanisms stricto sensu, other legal schemes could also be considered. For instance, some authors have advanced the merits of no-fault compensation schemes as a way to improve the victim’s position in light of AI’s disruptive characteristics.218 Such compensation regimes do not require the victim to prove a fault or even causation, but merely that the conditions for application are met. However, some have commented on the possible lack of deterrence such a mechanism could have on the parties made to compensate without fault.219 Such parties would usually have mandatory liability insurance in order to cover the costs of the accident. But no-fault insurance regimes can also have their costs covered by a privately or publicly managed fund.220 Insurance is another important component in the operationalization of tort law mechanisms that may be further explored.221 Comandé goes as far as to recommend using technologies to address the liability and accountability chal- lenges raised by AI.222 He argues that the technological and legal frameworks could be used in order to provide evidence of the decision-making process which could help determine the liabilities of involved stakeholders.223 However, the feasibility of such an approach has yet to be demonstrated.

218 See on no-fault compensation schemes for autonomous vehicles e.g., M. SCHELLEKENS, ‘No-Fault Compensation Schemes for Self-Driving Vehicles’, 10(2). Law, Innovation and Technology 2018, pp 319–320. S. VAN UYTSEL, ‘Different Liability Regimes for Autonomous Vehicles’, 2021, p 67–92; K. WATTS, ‘Potential of No-Fault Comprehensive Compensation Funds to Deal with Automation and Other 21st Century Transport Developments’, 12. European Journal of Commercial Contract Law 2020, pp 1–21; D. LEVY, ‘Intelligent No-Fault Insurance for Robots’, 1. Journal of Future Robot Life 2020, pp 35–57.

219 See e.g., O. RACHUM-TWAIG, ‘Whose robot is it anyway? Liability for artificial intelligence based products’, University of Illinois Law Review 2020, p 1176; J.M. ANDERSON et al., ‘The US experi- ence with no-fault automobile insurance: a retrospective’ (Rand Corporation: Santa Monica 2010), p 11–12.

220 S. VAN UYTSEL, ‘Different Liability Regimes for Autonomous Vehicles’ (2021), p 89. 221 See on the interaction between tort law and insurance e.g., R. MERKIN & J. STEELE, ‘Chapter 9.

Compulsory Liability Insurance’, in Insurance and the Law of Obligations (Oxford: Oxford University Press 2013), p 251 and following; G. WAGNER, ‘Tort Law and Liability Insurance’, The Geneva Papers on Risk and Insurance – Issues and Practice (Apr. 2006), pp 277–292; J. SPIER, ‘Tort Law and Liability Insurance Influences and Interactions Premiere Partie – Articles – Special File: European Group on Tort Law’, Revue Hellenique de Droit International 1999 (52), pp 71–86.

222 G. COMANDÉ, ‘Multilayered (Accountable) Liability for Artificial Intelligence’ (2019), pp 165–183. He suggests that ‘AI requires a gradual layered approach to liability grounded on accountability principals (already embedded in the EU legal systems). AI requires the use of technology itself to unfold a multilayered accountable liability system and solve the ‘splitting the bill problem’ (p 177).

223 G. COMANDÉ, ‘Multilayered (Accountable) Liability for Artificial Intelligence’, p 170.

298

Reproduced with permission of copyright owner. Further reproduction prohibited without permission.