PaperlayerView exact revision on arXiv

Source-linked AI summary

The European AI Liability Directives -- Critique of a Half-Hearted Approach and Lessons for the Future

Philipp Hacker

arXiv:2211.13960v6cs.CYcs.AIcs.LG

TL;DR

The article analyzes the Commission’s two AI liability proposals and develops amendments and a future regulatory agenda. It concludes that the proposals fall short of a coherent EU framework and recommends one harmonizing regulation.

  • Problem

    The Commission’s two liability proposals do not provide a uniform EU framework that balances ease of compensation with legal certainty for AI development and deployment.

    Source: §VI · p. 74

  • Method

    The article analyzes the proposals’ provisions and develops recommendations for a coherent liability framework and rules addressing high-risk use, innovation, fairness, explainability, and sustainability.

    Source: p. 2, §VI · p. 74

  • Results

    The proposals do not form a convincing whole, and the article argues that they unjustifiably equate high-risk AI under the AI Act with high-risk AI under liability law.

    Source: §VI · p. 74

  • Takeaways & Limitations

    EU lawmakers should replace the two disparate directives with one coherent regulation that harmonizes liability for products, including software and AI.

    Source: §VI · p. 74

Abstract

from arXiv · show

As ChatGPT et al. conquer the world, the optimal liability framework for AI systems remains an unsolved problem across the globe. In a much-anticipated move, the European Commission advanced two proposals outlining the European approach to AI liability in September 2022: a novel AI Liability Directive and a revision of the Product Liability Directive. They constitute the final cornerstone of EU AI regulation. Crucially, the liability proposals and the EU AI Act are inherently intertwined: the latter does not contain any individual rights of affected persons, and the former lack specific, substantive rules on AI development and deployment. Taken together, these acts may well trigger a Brussels Effect in AI regulation, with significant consequences for the US and beyond. This paper makes three novel contributions. First, it examines in detail the Commission proposals and shows that, while making steps in the right direction, they ultimately represent a half-hearted approach: if enacted as foreseen, AI liability in the EU will primarily rest on disclosure of evidence mechanisms and a set of narrowly defined presumptions concerning fault, defectiveness and causality. Hence, second, the article suggests amendments, which are collected in an Annex at the end of the paper. Third, based on an analysis of the key risks AI poses, the final part of the paper maps out a road for the future of AI liability and regulation, in the EU and beyond. This includes: a comprehensive framework for AI liability; provisions to support innovation; an extension to non-discrimination/algorithmic fairness, as well as explainable AI; and sustainability. I propose to jump-start sustainable AI regulation via sustainability impact assessments in the AI Act and sustainable design defects in the liability regime. In this way, the law may help spur not only fair AI and XAI, but potentially also sustainable AI (SAI).

I. Introduction

The proposals respond to urgent questions of responsibility as AI expands, but opacity can impede proof of fault, defectiveness, and causality. The paper surveys their aims and scope, then argues that the two-directive approach leaves fragmentation and regime-boundary problems.

I. Introduction

  • New generative AI models intensify the question of who should bear legal responsibility when AI systems make mistakes.The paper frames their growing range of tasks and outputs as part of the regulatory urgency.Source: §I · p. 9

II. Risks of AI

  • AI autonomy can produce outputs unforeseeable even to developers, while technical and institutional opacity can make proving fault, defectiveness, and causality onerous or nearly impossible.These risks shape the evidence problems the liability regime must address.Source: §II · p. 12

III. Overview of the new regime

  • The proposals primarily use disclosure rules and case-specific alleviations of the burden of proof, while retaining different obligations for high- and non-high-risk systems.The AILD and PLD proposals set up an intricate system across model uses and addressees.Source: §III · p. 14

1. The contentious choice of two directives

  • The two directives leave a dual track between EU product liability and largely unharmonized national tort law, while a single harmonized regime is politically unlikely.The paper notes that most Member States continue to cherish their national tort regimes.Source: §1 · pp. 14–15
  • The AILD targets AI while the revised PLD reaches software more broadly, so claims about using AI could send victims to the wrong compensation regime.The paper proposes aligning the AI regime with public claims about using AI to reduce friction and legal uncertainty for plaintiffs.Source: §1 · p. 15

2. General aim and thrust of the two directives

  • The Commission presents the proposals as promoting effective compensation, legal certainty, insurability, innovation, and trustworthy AI.The proposals also aim to minimize risks of digital products.Source: §2 · pp. 15–16

3. Relationship and key differences between the two directives

  • Although the directives apply in parallel to different, nominally mutually exclusive aspects, injured persons may proceed under either framework or both.Their damage coverage, addressees, and substantive liability requirements differ.Source: §3 · p. 16

4. The contentious choice of two directives

The paper argues that AI liability should be unified in one EU regulation, while criticizing the Commission proposals’ reliance on AI Act definitions and risk categories. In particular, liability should reflect significant risks to individuals even when aggregate social risk is low.

4. The contentious choice of two directives

  • A single EU regulation would reduce divergent transpositions and help consumers understand which regime applies when suing providers across Member States.The paper argues Article 114 TFEU should support a comprehensive EU-wide framework for AI liability and software.Source: §4 · p. 18

1. Applicability

  • The AILD’s reliance on AI Act concepts leaves its scope uncertain while that Act is still being revised, and the paper questions transplanting those concepts into liability rules.The proposals differ substantially in material and personal applicability.Source: a) AILD Proposal: transplants from the… · pp. 18–19, §1 · p. 18
  • The AI definition risks covering relatively simple statistical and rule-based software unless “autonomy” requires learning capacity or a complexity threshold.The paper identifies distinguishing AI from complex but traditional software as a core problem.Source: (1) Definition of AI · p. 19

(2) Risk categories

  • The AILD should decouple liability’s risk classification from the AI Act because its high-risk category can be both over- and underinclusive for individual harm.The AILD’s disclosure rules and parts of its burden-of-proof relief apply only to high-risk AI systems.Source: (2) Risk categories · p. 20
  • AVs, emotion recognition, and insurance pricing may pose significant risks to individuals even when aggregate risk does not qualify them as high-risk under the AI Act.The paper argues that low total expected damage can coexist with substantial risk concentrated on a few people, warranting high-risk treatment under the AILD.Source: (a) Under-inclusiveness: missing high- · p. 21

(b) Over-inclusiveness: ChatGPT et al.

The paper argues that high-risk rules should attach to general-purpose and foundation models only in specific high-risk deployment cases. It also describes the revised PLD’s broader software reach and its expanded range of potentially liable actors.

(b) Over-inclusiveness: ChatGPT et al.

  • The Council approach could subject nearly all general-purpose AI systems to high-risk obligations and stricter AILD and PLD liability, including widely used open-source tools.The Parliament’s staggered approach instead distinguishes minimum provider duties, high-risk-use obligations, and value-chain information sharing.Source: (b) Over-inclusiveness: ChatGPT et al · p. 22
  • The paper proposes applying high-risk regulation, including liability, to general-purpose and foundation models only in specific high-risk use cases for which they are deployed.This recommendation responds to the broad scope of high-risk obligations under the Council approach.Source: (b) Over-inclusiveness: ChatGPT et al · p. 22

ii. Personal applicability

  • The revised PLD expressly includes software and digital manufacturing files as products, extending product liability beyond the AI Act’s high-risk systems to non-high-risk AI and other software.The paper describes this as the long-awaited extension of product liability to the digital economy.Source: (1) Software and AI · p. 24, b) PLD Proposal: a new software… · p. 24
  • The revised PLD excludes free and open-source software developed or supplied outside commercial activity, an exemption the paper says should be placed in the product definition.The paper notes that contributors might otherwise be discouraged from contributing.Source: (2) Open-source software · p. 24
  • The PLD expands potential defendants to related-service providers, authorized representatives, fulfilment service providers, and, in certain circumstances, online platforms.The proposal also allows distributors to be sued if the manufacturer and other specified entities cannot be identified.Source: (1) Economic operators · p. 25
  • Digital services essential to a product’s function can fall within PLD liability, including AI software and providers of training data.The paper treats such services as related services and argues that training-data developers are subject to product liability.Source: (2) Embedded software and related… · p. 26

2. Substantial trigger of liability under the directives

The AILD uses fault as its liability trigger, with procedural relief generally tied to violations of AI-specific rules. The paper questions limiting that relief to cases where AI decisions proceed without substantive human intervention.

2. Substantial trigger of liability under the directives

  • Under the AILD, fault is the liability trigger, and disclosure and burden reversal generally require a substantive violation of the AI Act.In some cases, breach of an AI-specific duty of care outside the AI Act can suffice.Source: a) AILD Proposal: Fault · p. 27
  • The proposal withholds evidentiary alleviations when a human intervenes between AI output and damage, granting them only where decisions proceed fully automatically after the output.The stated rationale is that human intervention leaves causality and non-compliance as difficult to prove as in non-AI cases.Source: a) AILD Proposal: Fault · p. 27
  • A fully automated decision standard may invite circumvention if human actors merely rubberstamp AI outputs, while substantive human intervention provides a broader boundary.The paper compares this concern with the interpretation of automated decision-making under Article 22 GDPR.Source: a) AILD Proposal: Fault · p. 27

b) PLD Proposal: defectiveness of the product

The PLD Proposal expands the circumstances considered in assessing product defectiveness, but AI cases chiefly raise difficult design-defect questions. The paper proposes a system-level standard that accounts for technical standards, performance relative to the status quo, and reasonable alternative designs.

Learning after deployment

  • The PLD Proposal considers a system’s potential to learn after deployment, requiring operators to manage risks from adaptive behavior and user-provided data.The paper identifies curation or manual supervision as possible safeguards against deterioration from erroneous or discriminatory user interactions.Source: (1) Learning after deployment · p. 29

Control after deployment

  • Manufacturers remain liable for defects while they retain control of a product after market placement, including through partial remote control during operation.The rule applies whether or not the system continues learning.Source: (2) Control after deployment · p. 30

Product safety and cybersecurity

  • The PLD Proposal links defectiveness to product-safety requirements, including cybersecurity requirements drawn from existing and proposed EU digital legislation.The paper notes references to the GDPR, the proposed Cyber Resilience Act, and Article 15 of the AI Act.Source: (3) Product safety and cybersecurity · p. 30

Defects in AI systems

  • Most AI liability cases are expected to concern design defects, such as errors shared across comparable models rather than faults unique to one product copy.Examples include image recognition that misses structures, scoring systems that fail for particular income levels, and face recognition that misidentifies older people.Source: ii. Defects in AI systems · p. 31

System-oriented approach

  • The paper favors a system-oriented defect test over judging an AI error by whether an average human would make it, because the latter may over-deter and is difficult to operationalize.A system-oriented test assesses risks across the AI system’s deployments and focuses incentives on the modeling stage, where design is generally determined.Source: (2) AI-specific approach · pp. 31–32, (1) Anthropomorphic approach · p. 31

Technical standards

Infrahuman performance

  • Below-human-average performance should trigger a rebuttable presumption of defectiveness, assessed against the status quo and reasonable alternative designs.The paper also calls for warnings about errors that cannot be efficiently avoided and says operators should demonstrate safety improvements over a purely human process where relevant.Source: (b) Infrahuman performance · pp. 32–33

Suprahuman performance

  • Above-human performance is not a defect-free safe harbour: operators must show no reasonable alternative design, positive social utility, and warnings for unavoidable errors.These conditions aim to preserve incentives for improvement while protecting affected persons through alternative designs or warnings.Source: (c) Suprahuman performance · p. 34

3. Disclosure of evidence and access to information

The proposals use disclosure procedures to address claimants’ information disadvantage in AI liability cases, but the procedures differ in scope and access. The paper recommends harmonization and changes to make disclosure practically usable while protecting defendants’ legitimate interests.

Article 3 AILD Proposal

  • The AILD disclosure procedure lets courts order necessary, proportionate disclosure for high-risk AI claims, subject to plausibility, prior attempts to obtain evidence, and safeguards for confidential information.Noncompliance with a disclosure order creates a rebuttable presumption of non-compliance with the relevant duty of care.Source: a) Article 3 AILD Proposal · p. 35

Article 8 PLD Proposal

  • The PLD disclosure mechanism also covers non-high-risk AI systems, and noncompliance with an order creates a rebuttable presumption of defectiveness.Unlike the AILD procedure, a PLD claimant need not first ask the defendant to provide the evidence or make proportionate attempts to obtain it.Source: b) Article 8 PLD Proposal · p. 36

Pretrial access

  • AILD disclosure may be sought before or during trial, while the PLD procedure appears available only after a claim is filed, potentially increasing claimants’ costs.The paper recommends pretrial disclosure under the PLD to lower barriers to compensation.Source: b) Article 8 PLD Proposal · p. 36, c) Assessment · p. 36

Rationale and effects

  • Disclosure can help potential claimants decide whether to sue, support compliance incentives, and screen out claims that lack sufficient evidence.The paper notes that AILD disclosure largely overlaps with AI Act documentation and record-keeping obligations, so the added burden for high-risk providers may be low.Source: i. Merits · pp. 36–37

Shortcomings and recommendations

  • AI Act disclosures are designed for technically knowledgeable recipients, so liability proceedings need information that claimants and judges can understand without costly expert interpretation.The paper warns that otherwise disclosure risks having little practical effect despite the documentation already required by the AI Act.Source: ii. Shortcomings and recommendations · p. 37
  • To reduce barriers, the paper proposes triggering disclosure with evidence of damage and AI involvement, while retaining the plausibility threshold for competitor claims.It also calls for strict trade-secret safeguards in competitor cases.Source: ii. Shortcomings and recommendations · p. 37
  • Disclosure should reach information the defendant legally should possess and be available in the court where the main case can be brought.The paper argues that a defendant’s failure to keep legally required records should not eliminate the disclosure duty, and that remote venue requirements can impose prohibitive costs.Source: ii. Shortcomings and recommendations · p. 38

Harmonization of mechanisms

4. Burden of proof

Both proposals ease claimants’ burden of proof for selected elements of AI liability claims. Their rules address different underlying elements but also reverse proof concerning specified causal links.

Elements and causality

  • The AILD Proposal addresses proof of non-compliance with a duty of care, while the PLD Proposal addresses product defectiveness.The proposals also reverse the burden of proof for specified causal links involving fault and AI output, or defectiveness and damage.Source: §4 · p. 39

Scope of changes

  • The proposals’ targeted burden shifts concern fault or duty-of-care non-compliance, defectiveness, and causality.The paper notes that scholars and expert groups had proposed similar changes before the Commission proposals.Source: §4 · p. 39

Comparison of proposals

  • The two directives assign different liability elements to the proof rules while sharing a focus on easing proof of causal links.The AILD links breach of a duty of care to AI output; the PLD links defectiveness to damage.Source: §4 · p. 39

a) AILD Proposal – general elements

The AILD Proposal leaves key liability concepts to Member States rather than defining them. The passage also links opacity in some AI models to difficulty identifying which input features mattered for an output.

AILD Proposal – general elements

b) PLD Proposal – general elements

The proposals seek to ease proof of AI-related harm through disclosure and rebuttable presumptions, but claimants still face substantial hurdles in showing fault and causation. The presumptions apply only under specified conditions, and the user-focused rule has a notable post-output oversight gap.

b) PLD Proposal – general elements

  • Under the PLD Proposal, claimants generally must prove defectiveness, damage, and the causal link between them.Source: b) PLD Proposal – general elements · p. 41
  • Claimants may lack access to training data and still need to show that the data choice caused the AI system’s output, even when a balanced dataset might not have prevented the error.The proposals seek to provide redress by reversing the burden of proof in certain circumstances.Source: c) A concrete example: the bank robbery · p. 41
  • The AILD causality presumption requires proof of fault, a reasonable likelihood that fault influenced the output, and that the output caused damage, plus at least one specified scenario.Fault must involve breach of a duty of care intended to protect against the damage; the proposal also preserves the claimant’s duty to show that the AI output caused the damage.Source: d) AILD Proposal: non-compliance and… · p. 42, (1) Three necessary conditions · pp. 42–43, (2) Four individually sufficient… · p. 43
  • For high-risk provider claims, the presumption depends on breaches of specified AI Act requirements, yet claimants may need disclosure to establish those breaches.If the presumption applies, developers must provide expert testimony to rebut the causal link; claimants still bear the cost of proving the initial breach.Source: (2) Four individually sufficient… · pp. 43–44
  • For high-risk user claims, a breach of post-output human-oversight duties may not affect the already-produced output, undermining the causality presumption’s stated objective.The claimant would need to show that oversight would have corrected the output, rather than changed it.Source: (b) Case 2: claim against users of high- · pp. 44–45
  • For non-high-risk AI, the causality presumption relies on a general duty of care and applies only when the court finds proof excessively difficult.Source: (c) Case 3: claim concerning non-high- · p. 45

(d) Case 4: claim against consumers

The proposed consumer presumption is limited to certain cases of improper operation and does not cover routine use. The paper also identifies broader limits in the proposals, including gaps for post-output duties and the difficulty of proving malfunction in scoring systems.

(d) Case 4: claim against consumers

  • The causality presumption fails to capture duties performed after an AI output, including human oversight, because those actions cannot change the output already produced.The paper notes that transparency explanations may also arrive after output generation, creating a similar problem.Source: e) Strengths, weaknesses and… · p. 47
  • The consumer causality presumption excludes routine use and applies only to specified operational interference or failure to set required conditions when the consumer knew or should have known the system would not work properly.The consumer must have materially interfered with operating conditions or failed to determine them despite being required and able to do so.Source: (d) Case 4: claim against consumers · p. 46, e) Strengths, weaknesses and… · p. 47
  • A claimant can invoke the PLD’s defectiveness presumption for disclosure non-compliance, violation of mandatory safety requirements, or obvious malfunction, though the latter two grounds must be established by the claimant.Establishing the grounds may require costly expert analysis of data defendants disclose.Source: i. Sufficient condition for… · p. 49, g) Strengths, weaknesses and… · p. 51
  • The proposals lower enforcement barriers in obvious-malfunction cases, but claimants in other cases may need expensive experts and access to litigation funding.Source: g) Strengths, weaknesses and… · pp. 51–52
  • Obvious malfunction is easier to establish for classification errors than for continuous-score predictions, where accessible ground truth may not exist and expert analysis may be needed.The paper identifies credit scoring, insurance pricing, and job-performance prediction as examples of continuous-score tasks.Source: g) Strengths, weaknesses and… · pp. 51–52
  • The causality presumption should not be rebutted merely by asserting that a machine-learning model could have made the same error under alternative training.The paper argues that accepting that rebuttal would make the presumption ineffective in machine-learning cases.Source: i) Rebuttal of the presumptions · p. 52

5. Elements exclusive to the PLD Proposal

The PLD Proposal broadens eligible damages by removing the property-damage threshold and covering some data loss, but excludes pure economic loss. That exclusion leaves victims in several high-risk economic applications to pursue claims under the AILD Proposal.

5. Elements exclusive to the PLD Proposal

  • Removing the €500 property-damage threshold sets incentives to avoid defects even when they cause small losses, which may aggregate into substantial uncompensated harm.The author also argues that commercial property should be eligible because excluding it reflects an outdated view of product liability as consumer protection alone.Source: a) Expansion of eligible damages · pp. 52–53
  • The PLD Proposal adds loss or corruption of data not exclusively used for professional purposes to eligible damage.The author considers this necessary while data ownership remains unresolved.Source: a) Expansion of eligible damages · p. 53
  • The PLD Proposal covers only damage to property, life, or health, excluding pure economic loss common in employment, credit-scoring, and insurance-pricing cases.Victims in those settings are relegated to the AILD Proposal, where they must prove fault without the PLD’s defectiveness presumptions.Source: a) Expansion of eligible damages · p. 53
  • The author argues that pure economic loss should be eligible for high-risk AI claims or the regime may remain ineffective in the very applications it targets.This proposed change addresses the exclusion of direct financial losses from employment, credit-scoring, and insurance-pricing decisions.Source: a) Expansion of eligible damages · p. 53

c) Development-defect defence

The PLD Proposal retains the development-risk defence but extends manufacturers’ exposure when they retain control and a defect becomes discoverable. The author proposes a compensation fund as a compromise between innovation and injured persons’ compensation.

c) Development-defect defence

  • The PLD Proposal retains the development-risk defence, removes Member States’ option to derogate from it, and adopts an objective assessment of scientific and technical knowledge.Manufacturers remain liable if they retain control after market placement or service launch and the defect becomes discoverable during that period.Source: c) Development-defect defence · pp. 54–55
  • The defence may spur innovation, but it shifts losses to users who may be no better placed than manufacturers to understand complex defects.The author frames this as a continuing tension between innovation incentives and injured persons’ compensation.Source: c) Development-defect defence · p. 55
  • The author proposes an EU-wide fund to compensate victims who would have had a claim absent the defence, while preserving incentives for users’ diligence.The fund would compensate injured persons for development defects without making end users residual loss bearers; contributory negligence would still be factored in.Source: c) Development-defect defence · p. 55

V. The way forward

The author maps a future AI-liability framework around effective compensation, innovation, trustworthy AI, and sustainability. For liability, the proposed roadmap differentiates strict liability from rebuttable presumptions according to the model, deployment, and party involved.

V. The way forward

  • The roadmap identifies effective compensation, innovation and legal certainty, trustworthy AI, and sustainability as four areas for improving the proposed framework.The author notes that the article sketches possible scenarios and leaves fuller exposition to follow-up papers.Source: §V · p. 56

1. Effective compensation: strict liability versus negligence revisited

  • The author calls for a unified AI-liability framework because fault resurfaces in many PLD design- and warning-defect claims, making the supposed divide from AILD fault liability largely artificial.Claimants may need to prove a breach of duty under either route, and the current proposals do not achieve the proposed coherence.Source: a) Foundations: fault under the AILD… · pp. 56–57
  • Economic operators and professional users of high-risk AI would face strict liability based essentially on causation, while developers and all users of prohibited systems would also face it.The author says strict liability can influence activity levels and avoid costly inquiries into technical duties or defects.Source: b) Proposal for a comprehensive AI… · p. 58, (1) General reasons for truly strict… · p. 58
  • The author distinguishes models that should not cause harm when operating properly from models meant to cause specific harms, linking that distinction to strict liability or a presumption of fault.For legitimate-harm models, the author says defectiveness and breach of duty remain necessary to limit liability; clear defectiveness thresholds are therefore important.Source: (2) Restriction to illegitimate-harm… · p. 59, (3) Combination with pure economic loss · p. 60
  • Pure economic loss with strict liability would be limited to prohibited systems and high-risk systems that typically cause financial harm, alongside liability caps and SME insurance subsidies.The author says these safeguards allow strict liability and pure-economic-loss recovery to coexist; an insurance exemption for gross negligence is also proposed to counter moral hazard.Source: (3) Combination with pure economic loss · pp. 59–60
  • For legitimate-harm models, defectiveness, breach of duty, and causation would be presumed but rebuttable through disclosure of training data and fairness information.The author presents this as an information-forcing rule that can encourage disclosure by developers and users who have an information advantage.Source: (3) Combination with pure economic loss · pp. 60–61
  • Foundation Model liability would attach to specific high-risk deployments, with providers and system operators jointly and severally liable and able to allocate compliance duties internally.The proposal includes developers that fine-tune the model and allows internal indemnification depending on who was responsible for the source of harm.Source: iii. Foundation Models · p. 61
  • SMEs and non-high-risk AI would face rebuttable presumptions, consumers fault-based liability with a limited presumption of causality, and defendants could invoke contributory negligence.For SMEs, the presumption could apply during the first three years after market placement; the proposed consumer rule avoids a presumption of fault or strict liability.Source: b) Proposal for a comprehensive AI… · p. 57, iii. Foundation Models · p. 62

2. Innovation

The author proposes aligning liability duties with the AI Act where possible, while using quantitative safe harbours to improve legal certainty. Because the AI Act does not adequately address all compensable risks, compliance would not fully shield companies from civil liability.

2. Innovation

  • The author proposes aligning PLD and AILD duties with AI Act obligations, ideally making AI Act violations necessary and sufficient for liability.Full alignment would reduce compliance and enforcement costs, but the author says AI Act compliance cannot fully shield companies because some compensable risks remain unaddressed.Source: a) Alignment with the AI Act · pp. 62–63
  • Because full alignment is not desirable at present, the author recommends quantitative, use-case-specific safe harbours based on harmonized standards and specifications.Examples include converting broad requirements for training data, transparency, and accuracy into measurable thresholds.Source: b) Safe harbours and legal certainty · p. 63, a) Alignment with the AI Act · p. 63
  • Quantitative compliance and non-compliance thresholds could mark green, grey, and red zones, giving developers and users greater legal certainty.The author links legal certainty to reduced chilling effects and says developers below safe-harbour thresholds would deploy at their own risk.Source: b) Safe harbours and legal certainty · pp. 63–64

3. Trustworthy AI

The paper treats trustworthy AI as requiring liability to address risks including unforeseeable behavior, opacity, and discrimination. It argues that the proposals should better address harm from unforeseeable AI outputs.

3. Trustworthy AI

  • Trustworthy AI regulation should address unforeseeable behavior, opacity, and discrimination, alongside compensation and innovation.The paper identifies hedging unforeseeable behavior, reducing opacity, and mitigating discrimination as key aspects of trustworthiness.Source: §3 · p. 64
  • Non-manipulation and privacy or data protection are identified as trustworthy-AI concerns beyond this paper’s scope.Source: §3 · p. 65
  • AI Act requirements for representative data and lower error rates are presented as conditions that can reduce unforeseeable damage.The paper points to data governance intended to support generalization and output requirements intended to minimize errors and prevent negative feedback loops.Source: a) Autonomy and unforeseeability · p. 65
  • Because causality is not harmonized and some national systems exclude objectively unforeseeable damage, both directives should cover harm caused by unforeseeable AI acts or omissions.The paper argues that AI damage can be foreseeable in the abstract even when a particular output is unforeseeable.Source: a) Autonomy and unforeseeability · p. 65

b) Complexity and opacity: explainable AI

Disclosure mechanisms target institutional opacity, while technical opacity presents a distinct challenge. The paper argues that liability rules need clearer causal presumptions and stronger coverage of explainability and non-discrimination.

b) Complexity and opacity: explainable AI

  • Disclosure and burden-of-proof rules target institutional opacity, but technical opacity arises from model complexity and is less explicitly addressed.Institutional opacity involves withholding information; technical opacity concerns difficulty identifying the causes of model outputs.Source: b) Complexity and opacity: explainable… · p. 66
  • For complex models, explanations may be post-hoc approximations, global explanations are harder, and their accuracy remains debated.The passage contrasts these limits with linear regression and simple decision trees, which can provide feature weights more readily.Source: b) Complexity and opacity: explainable… · p. 66

ii. Explainable AI: Presumptions of causality

ii. Discrimination and the PLD Proposal

  • Under the PLD, non-discrimination should inform defectiveness case by case, with implementing acts clarifying acceptable fairness metrics for particular uses.The passage notes that fairness metrics can be incompatible or only partially satisfiable.Source: ii. Discrimination and the PLD Proposal · p. 70

4. Sustainability

The paper argues that AI regulation should address rising sustainability risks as well as liability, innovation, explainability, and fairness. It proposes sustainability assessments and sustainable design defects while calling for a more coherent EU liability framework.

4. Sustainability

  • The paper describes AI sustainability as a major social and economic concern and notes that machine-learning emissions are real, noticeable, and rising.It reports an estimated ICT contribution of 1.8–3.9% of global greenhouse-gas emissions, potentially exceeding global air traffic at about 2.5%.Source: §4 · p. 71, a) A blind spot in AI regulation · p. 71

b) Regulating for sustainable AI

ii. Sustainable design defect

  • A model with suboptimal sustainability could qualify as a design defect if a feasible, comparably performing alternative with a smaller carbon footprint was not realized.Source: ii. Sustainable design defect · p. 73

VI. Conclusion

  • The proposals take steps in the right direction but do not provide a uniform, coherent EU framework; the paper calls for one fully harmonizing regulation.It also argues that the proposals collapse distinctions between social and individual risk by equating high-risk AI under the AI Act with liability categories.Source: §VI · p. 74
  • The proposed future framework combines staggered liability, innovation support, and provisions for explainable and non-discriminatory AI.Its liability levels range from strict liability for certain high-risk applications to presumptions for SMEs and non-high-risk AI, and fault-based liability for consumers.Source: §VI · p. 74
  • The EU could spur a “Green Brussels Effect” if it acts first on sustainable AI, potentially advancing green technology beyond Europe.Source: §VI · p. 75

1. General considerations

The recommendations span applicability, proof, liability, and eligible damages. The author proposes targeted rules for high-risk AI and amendments to address black-box systems, discrimination, explainability, and sustainability.

1. General considerations

  • General-purpose AI and Foundation Models should face high-risk liability only in specific high-risk use cases, while consumers generally remain subject to regular fault.The author also recommends a unified regulation rather than two directives and rejects the AILD/PLD dichotomy.Source: §1 · p. 76

2. Applicability

  • High-risk rules should cover Foundation Models only for specific high-risk uses; open-source exceptions and responsibility allocation should be clarified across the AI Act and liability proposals.The author also recommends including autonomous vehicles, emotion recognition, and insurance pricing in the high-risk category, or extending AILD rules to them if that fails.Source: §2 · p. 76

3. Defectiveness and breach of duty

  • Technical standards should presumptively define defectiveness, and the AILD non-compliance presumption should apply even when a human makes the final decision.The author also recommends deleting the AILD recital reference to automated decision-making.Source: §3 · p. 76

4. Disclosure of evidence

  • Disclosure should trigger on evidence of damage and AI involvement, reach information legally at defendants’ disposal, and be harmonized across the proposals.Recommendations also include preservation measures, pretrial discovery, accessible formats, and aligned court procedures.Source: §4 · p. 77

5. Burden of proof

  • Claimants should be able to invoke defectiveness and causality presumptions together, with funding support for expert analysis and aligned AILD/PLD likelihood conditions.The author also proposes extending disclosure to proof of mandatory safety requirements and adding presumptions for post-processing violations.Source: §5 · p. 77

6. Representative actions; 7. Defences

  • Representative actions should cover natural persons acting professionally, and the PLD should add an EU-wide fund for injuries from development defects.Source: §6 · p. 78, §7 · p. 78

8. Pure economic loss and eligible damage

  • For high-risk AI, eligible PLD damages should include pure economic loss and commercially used property, not only privately used property.Source: §8 · p. 78

9. The way forward

  • The future framework should place black-box proof consequences on developers and users and include harm from unforeseeable AI acts or omissions.AI Act non-compliance should be sufficient but not necessary for defectiveness or breach; standards should define compliance zones, while rules specify explanation and fairness requirements and add sustainability assessments and design defects.Source: ii. Explainability · p. 79, i. Unforeseeability · p. 79, b) Innovation · p. 79
Loading 2211.13960v6…