Allocating Responsibility in Autonomous AI Systems: A Tiered Governance Model Under EU Regulation
Open accessThe DOI matches its Crossref record; 15 of the 15 references it cites were found in a registry or an index; 15 of the 55 sources its claims cite were checked for support.
Abstract
Autonomous artificial intelligence (AI) systems increasingly participate in decision-making processes that affect individuals, markets, and public administration. Their growing autonomy complicates the attribution of legal responsibility, particularly within regulatory frameworks that were designed around identifiable human actors and relatively stable products. Although European Union instruments such as the GDPR, the AI Act, and the revised Product Liability Directive address specific dimensions of risk and compliance, they do not fully resolve how responsibility should be allocated across the lifecycle of complex AI systems. The difficulty does not lie so much in the absence of legal rules. Rather, it reflects the structural tension between traditional liability models and the distributed architecture of contemporary AI development and deployment. By examining how existing EU regulatory instruments interact, the paper identifies fragmentation in responsibility allocation that may weaken institutional accountability. It then proposes a tiered model of legal responsibility based on meaningful control at different stages of system design, deployment, and operational oversight. Rather than introducing new forms of legal personhood, the model seeks to clarify how existing doctrines can be interpreted and coordinated in order to maintain regulatory coherence and socially intelligible accountability in digitally mediated environments. The model allocates responsibility according to meaningful control within distributed systems, offering a structurally coherent alternative for EU governance.
Overview
CoChat Summary
TL;DR
The paper argues that EU responsibility allocation for autonomous AI is fragmented across the GDPR, AI Act, and revised Product Liability Directive, and that existing instruments do not coherently assign liability across the AI lifecycle. It proposes a tiered model that allocates responsibility according to meaningful control at the structural, operational, and systemic levels, while rejecting AI legal personhood.
Key findings
- EU frameworks overlap but allocate accountability differently: product liability addresses compensation and defect attribution, the AI Act imposes ex ante compliance and risk-management obligations, and the GDPR regulates accountability through individual rights and controller obligations. The same harmful incident may therefore trigger overlapping and potentially inconsistent accountability, as illustrated by autonomous vehicles intersecting product liability, AI Act, and GDPR rules.
- The proposed AI Liability Directive introduced a rebuttable presumption of causality for certain high-risk AI cases, but it did not resolve how responsibility should be allocated among multiple actors and was withdrawn in 2025. The paper states that this outcome underscores the limits of narrowly procedural solutions and that lowering evidentiary thresholds shifts rather than eliminates the evidentiary burden.
- The tiered model has three levels: structural responsibility for developers/providers (design, transparency, testing, documentation, safeguards, linked to AI Act Articles 11–16 and Article 9); operational responsibility for deployers/end users (contextual adequacy, monitoring, linked to AI Act Articles 26 and 72, with no escape through purely formal human sign-off); and systemic accountability through traceability and auditable design (linked to AI Act Articles 12–13 and CJEU case C-203/22).
- The paper rejects granting AI systems legal personhood because they lack moral agency, asset ownership, independent economic existence, and capacity to respond to legal sanctions; responsibility should remain human-centered.
Methods
- Doctrinal and analytical legal research methodology examining primary EU legal instruments, including the GDPR, the AI Act, and the revised Product Liability Directive, as well as academic literature on AI governance and liability.
- The analysis uses interpretative examination of legislative texts, regulatory frameworks, and scholarly debate, developing a conceptual model through normative reasoning and systematic comparison rather than empirical data.
- No empirical data were collected or analyzed; the study is qualitative and conceptual.
Limitations
- Stated in the paper: no new data were created or analyzed, and the study does not employ empirical data; it develops a conceptual model through normative reasoning.
- Stated in the paper: the AI Act remains a regulatory compliance framework rather than a civil liability regime, so compliance does not automatically resolve liability questions; the proposed AI Liability Directive was withdrawn in 2025.
- Evident from the text: the tiered model is not empirically tested or implemented, and the paper notes that the distribution of AI Act obligations across providers and deployers corresponds only “broadly, though not perfectly” to the model’s structural and operational dimensions.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
14 Valid · 2 Not valid · 31 Not checked · 71 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“AI systems challenge this assumption because their behavior may evolve after deployment through continuous learning, software modifications, or changing operational inputs (Buiten 2024).”
What the sources say
- Supportedcites Buiten 2024 source
Section 4.1 says an AI system may continue learning after deployment, while section 5.1 discusses software updates after market introduction and the resulting change to when defectiveness is assessed. Section 4.3 also explains that performance depends on how closely the operating environment resembles the conditions used in testing. Together, these passages support the claim that AI behavior need not remain fixed after deployment. (LLM, conf=0.88)
hedgedstates a possible cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Harmful outcomes may involve developers, software providers, deployers, or other actors simultaneously, without a clear mechanism for allocating responsibility internally among them (Wendehorst 2020; Buiten 2024; Ebers 2020).”
What the sources say
- Partly supportedcites Wendehorst 2020 sourceExperimental — accuracy not yet measured
“A The possible addressees” (page unavailable) identifies producers, operators and, in some circumstances, other providers as potential parties bearing liability. The discussion of IoT complexity describes harm that could stem from a device, a separately purchased sensor or another provider’s data. But the paper does not establish that these actors are simultaneously responsible or address how liability would be allocated internally among them; it instead proposes rules for identifying a liable operator. (LLM, conf=0.84)
- Partly supportedcites Buiten 2024 sourceExperimental — accuracy not yet measured
The Conclusion (page unavailable in our copy) says multiple stakeholders can make liability for AI-related defects difficult to determine. The Introduction and §4.1 discuss allocating responsibility between producers and users, but the paper does not establish the claimed absence of a clear mechanism for internal allocation among developers, software providers, deployers, and other actors. (LLM, conf=0.82)
- Not checked — no full text of the sourcecites Ebers 2020 source
hedged- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“This becomes particularly relevant in AI systems, where harmful outcomes may emerge without a single identifiable defect in the traditional sense (Buiten, 2024).”
What the sources say
- Supportedcites Buiten, 2024 source
Section 5.2 (page unavailable) says that identifying and proving defects in AI systems is challenging because their components are interconnected and it can be unclear whether a particular outcome counts as a defect. Section 4.2 (page unavailable) contrasts these problems with traditional manufacturing defects, supporting the paraphrase that harm need not trace to one readily identifiable defect in the traditional sense. (LLM, conf=0.89)
hedged- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Wachter et al. (2017) note that many automated systems do not clearly show whether meaningful human review has genuinely taken place.”
What the sources say
- Partly supportedcites Wachter et al. 2017 sourceExperimental — accuracy not yet measured
The section on limitations of Articles 15 and 22 says the required level of human involvement is unclear in practice and that nominal involvement may be enough to take a decision outside the rules on solely automated processing. The section proposing clarification of Article 22(1) likewise questions whether a human who does not intervene counts. These passages support uncertainty about what constitutes meaningful human review, but do not establish that many systems fail to show whether such review actually occurred. Page numbers are not available in our copy. (LLM, conf=0.83)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Recent institutional analysis of AI civil liability in the EU also confirms that questions of compensation, legal certainty, and risk allocation remain central after the adoption of the AI Act and the revised Product Liability Directive (Bertolini 2025).”
What the sources say
- Supportedcites Bertolini 2025 doi:10.2861/0075079
Section 3.2 (p. 75) says the liability framework remained incomplete despite the AI Act and the proposed PLD reform, and identifies liability apportionment and uncertainty as continuing problems. Section 4.2 says the revised PLD will not resolve all concerns about compensation for AI-related damage; the study’s concluding discussion (p. 125) links clear liability rules to legal certainty and businesses’ ability to manage risk. (LLM, conf=0.90)
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
64 more claims · 5 with their cited sources checked
Create a free account to see all 118 claims and every source verdict.
- Probably not a substantial claim. 29
- It may state two separate claims as one. 32
- The quote only partly states this claim. 1
Citations
Cites
The EU General Data Protection Regulation (GDPR)
Voigt, P., & von dem Bussche, A. (2017) · doi:10.1007/978-3-319-57959-7 · Cited by 3,001 · cited 1×
“…22 GDPR grants individuals the right not to be subject to a decision based solely on automated processing, including profiling, where such decisions produce legal effects or similarly significant effects (European Union 2016, Regulation (EU) 2016/679, Article 22; Voigt and von dem Bussche 2017).”
Wachter, S., Mittelstadt, B., & Floridi, L. (2017) · International Data Privacy Law · doi:10.1093/idpl/ipx005 · Cited by 1,308
Veale, M., & Zuiderveen Borgesius, F. (2021) · Computer Law Review International · doi:10.9785/cri-2021-220402 · Cited by 805 · cited 1×
“The risk-based structure of the AI Act, which classifies systems according to potential harm and calibrates regulatory obligations accordingly, has been described as a significant shift toward horizontal AI governance (Veale and Zuiderveen Borgesius 2021).”
Regulating Artificial Intelligence Systems: Risks, Challenges, Competencies, and Strategies
Scherer, M. U. (2015) · doi:10.2139/ssrn.2609777 · Cited by 563
Regulating Artificial Intelligence
Wischmeyer, T., & Rademacher, T. (2019) · Springer eBooks · doi:10.1007/978-3-030-32361-5 · Cited by 198
The European AI liability directives: Critique of a half-hearted approach and lessons for the future
Hacker, P. (2023) · Computer Law & Security Review · doi:10.1016/j.clsr.2023.105871 · Cited by 124
Laux, J., Wachter, S., & Mittelstadt, B. (2024) · Computer Law & Security Review · doi:10.1016/j.clsr.2024.105957 · Cited by 110
We, the Robots?: Regulating Artificial Intelligence and the Limits of the Law
Chesterman, S. (2021) · doi:10.1017/9781009047081 · Cited by 84
AI Systems Under Criminal Law: a Legal Analysis and a Regulatory Perspective
Lagioia, F., & Sartor, G. (2020) · Philosophy & Technology · doi:10.1007/s13347-019-00362-x · Cited by 75 · cited 1×
“In the absence of these elements, assigning legal responsibility directly to AI systems would either prove ineffective in practice or merely obscure the fact that liability ultimately remains with the human actors behind them (Chesterman 2021; Lagioia and Sartor 2020).”
Strict liability for AI and other emerging technologies
Wendehorst, C. (2020) · Journal of European Tort Law · doi:10.1515/jetl-2020-0140 · Cited by 70 · cited 2×
“Harmful outcomes may involve developers, software providers, deployers, or other actors simultaneously, without a clear mechanism for allocating responsibility internally among them (Wendehorst 2020; Buiten 2024; Ebers 2020).”
Research works printed but not linked: 7
- European Union (1985). Council Directive 85/374/EEC of 25 July 1985 on the approximation of the laws, regulations and administrative provisions of the Member States concerning liability for defective products. Official Journal of the European Communities L, 210, 29–33.
- (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 (General Data Protection Regulation). Official Journal of the European Union L, 119, 1-88.
- European Union (2024a). Directive (EU) 2024/2853 of the European Parliament and of the Council of 23 October 2024 on liability for defective products and repealing Council Directive 85/374/EEC. Official Journal of the European Union L, Available online: https://eur-lex.europa.eu/eli/dir/2024/2853/oj/eng.
- European Union (2024b). Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union L, Available online: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng.
- Wischmeyer, Thomas, and Rademacher, Timo (2020). Artificial Intelligence and Legal Personality. Regulating Artificial Intelligence, Springer. · DOI
- Wachter, 2024, Limitations and loopholes in the EU AI Act and AI Liability Directives, Yale Journal of Law & Technology, 26, 671
- Wischmeyer, Thomas, and Rademacher, Timo (2020). Regulating Artificial Intelligence, Springer. · DOI
Cited by
Cheong, B. C. (2026) · Cambridge Forum on AI Law and Governance · doi:10.1017/cfl.2026.10054 · Cited by 2
“The problem of attributing legal responsibility when the entity causing harm is structurally distinct from its human creators – whether through corporate form, avatar intermediation or autonomous AI operation – is a recurring challenge in technology law (Cheong, 2024b; Papastergiou et al., 2026).”
Metadata
- Authors
- Papastergiou, F., Quintero, B., & Marin, V.
- Affiliations
- University of Córdoba (ES)
- Department of Education, Faculty of Education Sciences and Psychology, University of Córdoba, 14071 Córdoba, Spain
- Corresponding author
- Papastergiou, Foteini
- Venue
- Social Sciences
- Publisher
- MDPI AG
- Published
- 2026-06-16
- Volume
- 15
- Issue
- 6
- Pages
- 392
- ISSN
- 2076-0760 (ISSN-L)
- OpenAlex
- W7164897303
- Cited by
- 1
- References
- 12
- Citations by year
- 2026: 1
- FWCI
- 33.32
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Social Sciences (publishedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
- Zenodo (CERN European Organization for Nuclear Research) (cc-by)
- Topics
- Social Sciences › Law › Digitalization, Law, and Regulation
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Physical Sciences › Computer Science › Artificial Intelligence › Law, AI, and Intellectual Property
- Keywords
- autonomous AI systems, legal responsibility, EU AI regulation, GDPR, EU AI Act, Product Liability Directive, accountability, distributed systems, regulatory coherence, AI lifecycle, legal personality
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-05, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-05 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.