1 Introduction

1.1 Background and motivation#s. 1

With the adoption of the Artificial Intelligence Act (hereafter “AI Act”, “the Act”, or “AIA”) on 13 June 2024, the European Union (EU) became the first jurisdiction to enact a comprehensive horizontal regulatory framework for Artificial Intelligence (AI) systems.1 It represents the EU’s attempt to strike a balance between the often conflicting objectives of 1) the protection of fundamental rights and public safety, and 2) the promotion of innovation and economic competitiveness, including legal certainty for providers and deployers of AI systems.2 This balancing act is central in the provisions governing the classification of high-risk AI systems.

The AI Act adopts a risk-based regulatory approach, under which the intensity of regulatory requirements aims to be proportional to the level of risk that an AI system poses to fundamental rights, health, and safety.3 While few AI uses are outright prohibited, the Act introduces a broad “high-risk” category triggering extensive obligations. While the risk-based architecture of the Act has attracted considerable scholarly attention, most contributions have concentrated on the classification framework at a macro level. Early work has highlighted the practical enforceability challenges that flow from the Act's layered governance structure, including the potential of dual-use and the ease of disguising the “true” market of an AI system.4 Subsequent contributions have examined the conceptual coherence of the Act's risk tiers and the assumptions underpinning the high-risk category,5 the Act's interplay with existing product-safety regulation,6 and the adequacy of its conformity-assessment mechanisms.7

Article 6(3) of the Act includes a mechanism to exempt individual AI systems that pose no significant risk from high-risk obligations, even if they operate in high-risk areas. The wording of the mechanism was finalised late in the legislative process,8 and has received limited doctrinal scrutiny. This thesis aims to complement the existing literature by examining the AI Act Article 6(3)(a) and (d) carve-outs, asking what it means, in operational terms, for an AI system to perform a “narrow procedural task” or a “preparatory task to an assessment.”9

1.2 The Article 6(3) proportionality “safety valve”#s. 2

Casting the regulatory net widely in sensitive domains ensures that AI systems operating in contexts where fundamental rights are at stake are subject to scrutiny. At the same time, this creates a risk of over-inclusion. Not every AI system deployed within a high-risk use context will, by virtue of its specific function, pose a significant risk of harm. Gradually recognising this, the legislator introduced Article 6(3) as what may be characterised as a proportionality “safety valve”: a mechanism that allows certain AI systems within Annex III domains to not be considered to be high-risk where they do “not pose a significant risk of harm to the health, safety, or fundamental rights of natural persons, including by not materially influencing the outcome of decision-making.”10 Article 6(3) sets out a closed list of four conditions under which the carve-out may apply, in direct quote:

(a)the AI system is intended to perform a narrow procedural task;

(b)the AI system is intended to improve the result of a previously completed human activity;

(c)the AI system is intended to detect decision-making patterns or deviations from prior decision-making patterns and is not meant to replace or influence the previously completed human assessment, without proper human review; or

(d)the AI system is intended to perform a preparatory task to an assessment relevant for the purposes of the use cases listed in Annex III,

with the following subparagraph establishing a carve-out from the carve-out in that an AI system in Annex III “shall always be considered to be high-risk where the AI system performs profiling of natural persons.”

Of the four conditions, (b) and (c) share the trait that they describe AI system activity that is independent from and temporally placed after a previously completed human activity or assessment. While (b) and (c) are not free of ambiguity,11 the backwards looking nature and independence from the previous human activity/assessment are concrete signposts for AI providers and deployers assessing if their system qualifies for exemption. On the other hand, the concepts of (a) “narrow procedural task” and (d) “preparatory task to an assessment” are open-textured and lack precise legal definition in the operative provisions of the Act. This is the motivation for narrowing the scope of this thesis to Article 6(3) conditions (a) and (d).12

This interpretative ambiguity gives rise to a dual risk. On the one hand, risk-seeking actors with an incentive to minimise regulatory costs may strategically exploit the carve-outs by characterising the functions performed by their AI systems in narrow terms, avoiding high-risk classification and compliance obligations. This risk of regulatory arbitrage has been identified as a structural concern in the literature on the AI Act.13 Risk-averse actors in heavily regulated sectors may, conversely, avoid the carve-outs altogether, preferring the legal certainty of full high-risk compliance – at the cost of unnecessary compliance burdens, dampened incentives for beneficial AI adoption, and a weakened proportionality principle.14

This thesis seeks to derive, through systematic legal interpretation, a set of operational criteria for whether an AI system’s task qualifies as “narrow procedural” or “preparatory to an assessment”. The criteria aim to reduce the scope for regulatory arbitrage by risk-seeking actors and to give risk-averse actors the legal certainty needed to invoke the carve-outs where they legitimately apply.15

1.3 Research question#s. 3

The open-textured conditions for exemption from high-risk classification require operationalisation if they are to serve their intended function as a proportionate corrective to the breadth of Annex III. This thesis addresses the following research question:

What criteria can be derived from the AI Act and analogies from relevant EU law to operationalise the notions of “narrow procedural task” and “preparatory task to an assessment” in Article 6(3)(a) and (d), and how can these criteria be applied in practice to reduce regulatory arbitrage while enabling legitimate use of the Article 6(3) carve-outs?

The research question contains two analytically distinct but interrelated components. The first is interpretative: it asks what substantive criteria can be derived from the legal sources to give operative content to the concepts of “narrow procedural task” under Article 6(3)(a) and “preparatory task to an assessment” under Article 6(3)(d). The second component is applied: it asks how such criteria can be operationalised to guide the practical classification of AI systems, to reduce the scope for regulatory arbitrage by risk-seeking actors while preserving the legal certainty necessary for risk-averse actors to make legitimate use of the carve-outs.

The scope of the inquiry is delimited in several respects. First, the thesis focuses on the conditions in Article 6(3)(a) and (d), rather than on all four alternative conditions listed in Article 6(3). As discussed in the previous section, the conditions of Article 6(3)(b) and (c), concerning AI systems that improve the result of a previously completed human activity and AI systems that detect decision-making patterns, are relatively easy to operationalise through their clear temporal and functional separation from decision processes. Second, the thesis does not undertake an empirical study of how providers and deployers currently intend to apply the Article 6(3) exemptions; the practical examples interspersed throughout the text and detailed in the appendix are illustrative and designed to test the derived criteria, not to make claims about industry practice.16

By answering this research question, I aim to provide tools that are of value not only for academic discourse, but also for practitioners, including AI providers, deployers, and national market surveillance authorities, who must navigate the Article 6(3) assessments in practice.

1.4 Methodology and sources#s. 4

For the interpretative part of the research question, the thesis employs legal dogmatic method: the provisions are interpreted through recognised EU-law sources and the ordinary textual, systematic, and teleological techniques. Because all language versions of EU legislation are equally authentic, a full multilingual comparison is relevant, but for reasons of scope, this thesis proceeds from the English text and uses other sources to control its interpretation.17

The analysis first examines the wording of Article 6(3)(a) and (d), especially “procedural”, “preparatory”, and “task”. It then reads those terms in the context of the Act as a whole, including Annex III, Article 14 on human oversight, Article 3 definitions, and the Act’s risk-based structure. Finally, it considers the purpose of Article 6(3): maintaining fundamental-rights protection while avoiding disproportionate high-risk classification for genuinely low-risk functions. GDPR Article 22 is drawn on by analogy to illuminate the boundary between preparatory processing and decision-influencing automation.18

The doctrinal analysis is supplemented by the syntactic/semantic distinction from information theory. The distinction is used only as an interpretive aid: it helps clarify whether an AI task operates on the form of information or its evaluative meaning, without displacing the legaldogmatic analysis.

The thesis is primarily concerned with lex lata. Section 5.3 adds de lege ferenda observations where the doctrinal analysis reveals structural weaknesses and unfortunate disincentives in the current carve-out framework.

In accordance with settled ECJ case law, the Act’s recitals are used as an aid to interpretation of the operative provisions, without being treated as autonomous sources of binding legal obligations.19 Recital (53) is particularly important because it gives examples of “narrow procedural” and “preparatory” tasks and explains the rationale of the exemption mechanism.

ECJ case law supplies both general interpretative principles and specific analogies. Because the Act has not yet generated relevant jurisprudence, SCHUFA is especially useful: its treatment of GDPR Article 22 shows that a functionally preparatory automated output may still be legally decisive where the subsequent decision strongly draws on it. The Court’s approach to derogations also frames Article 6(3): exceptions are interpreted strictly, but not so strictly as to deprive them of effet utile.20

Soft law is used only subsidiarily as interpretative aids. Article 6(5) requires Commission guidelines on the practical implementation of Article 6, but their absence at the time of writing leaves this interpretation without an authoritative soft-law benchmark.21

1.5 Thesis structure#s. 6

Chapter 2 situates Article 6(3) within the AI Act’s risk-based architecture, the provider/deployer allocation of responsibility, and Article 14 on human oversight. Chapter 3 interprets Article 6(3)(a) and (d). Chapter 4 turns that interpretation into six operational criteria. Chapter 5 answers the research question, summarises the contribution, and identifies points for legislative or supervisory clarification and improvement. The Appendix sets out the criterion-by-criterion application of the framework to three illustrative case studies summarised in Section 4.3.