MarIus

3.3.2.1 The starting point: a matter of interpretation

To determine whether ANS must be classified as "high-risk AI systems" under Articles 6(1) or (2), an interpretation of the provisions and their respective conditions is necessary. Interpretations of the AIA must be carried out in accordance with the method set out by the CJEU, being the supreme authority in all EU legal matters.^{(3) Treaty on European Union (TEU), (1992) OJ C 191/1, Article 19;Lorna Woods, Philippa Watson and Marios Costa, Steiner & Woods EU law (13th edn, Oxford University Press 2017) 46.} This essentially involves starting with a natural understanding of the wording in the context in which it stands, and in light of the objectives that the provision is intended to achieve.^{(4) Sorina Doroga and Alexandra Mercescu, 'A Call to Impossibility: The Methodology of Interpretation at the European Court of Justice and the PSPP Ruling' (2021) 13 EJLS 87, 95-96.} The AIA Article 1(1) sets out the "purpose" of the Act, and is therefore of great interest when interpreting Article 6. In contrast to Article 1(1), the Recitals and Explanatory Memorandum of the AIA are not binding, but they are still relevant, as they provide context.

3.3.2.2 Significance of the pending clarifications from the European Commission

However, conducting interpretations with a goal to find de lege lata in questions of high-risk classification may be premature as of April 2025. According to the AIA Article 6(5), the European Commission has an obligation to consult the European Artificial Intelligence Board, and then "provide guidelines specifying the practical implementation" of Article 6, as well as providing "a comprehensive list of practical examples of use cases of AI systems that are high-risk and not high-risk". The Commission has until 2 February 2026 to do so.

Given the broad language of Article 6(5), it can be assumed that the Commission has the authority to clarify whether ANS are "high-risk" or not. It is less certain whether the Commission is obligated to provide this clarification. Nonetheless, the primary concern for stakeholders in autonomous shipping must naturally be whether the Commission does or does not clarify this. Although it is difficult to predict with certainty what the Commission will address, it seems unlikely that it will bypass the practical question of ANS.

Article 6(5) does not specify whether the guidelines and the list of practical examples shall be considered legally binding for later interpretation. However, the European Commission has already approved guidelines on how to correctly understand both the concepts of "AI systems" and "prohibited AI systems".^{(5) Pursuant to the AIA Articles 96(1) f) and b), respectively.} Both sets of guidelines explicitly state that they are not to be considered binding, and that they do not affect the authority of the CJEU.^{(6) The Guidelines on the Definition of an AI system para. 7; The European Commission, 'Approval of the content of the draft Communication from the Commission – Commission Guidelines on prohibited artificial intelligence system practices established by Regulation (EU) 2024/1689 (AI Act)' COM (04 February 2025) 884 final (The Guidelines on Prohibited Artificial Intelligence Practices) para. 5.} Accordingly, it is unlikely that the guidelines and the list regarding high-risk classification will be considered binding either. However, the very delegation of authority to issue the guidelines and the list implies that they should carry decisive weight when interpreting Article 6 (1) and (2), unless superseded by the CJEU or other EU law. Therefore, clarifications from the Commission on whether ANS are high-risk AI systems will likely be of significant practical importance.

3.3.2.3 Legal relevance of current interpretations

Despite the pending clarifications from the European Commission, current interpretations of the AIA Articles 6(1) and (2) are not necessarily without value. On the contrary, they may even offer insights into what the Commission can – and arguably should – clarify.

Harmony within EU law is important when adopting and understanding new EU legislation. This has a bearing on both the principle of creating legal

certainty in EU law, and the function of removing barriers on the internal market through successfully harmonizing rules of law within the Union.^{(7) Woods, Watson, and Costa, Steiner & Woods EU Law, pages 171 and 335, respectively.} This suggests that the Commission should recognize the benefit of their understandings of the AIA upholding conformity with and indirectly promoting the interpretive method of EU law, as set out by the CJEU. Furthermore, though the CJEU will likely respect the clarifications, it still has the authority to overturn them, as recognized by the Commission in the previously adopted guidelines.^{(8) The Guidelines on the Definition of an AI system para. 7, The Guidelines on Prohibited Artificial Intelligence Practises para. 5.} This also suggests that the Commission should set out understandings that could have emerged from plausible interpretations of the AIA.

In the following sections, I interpret Articles 6(1) and (2), considering whether ANS can be classified as "high-risk AI systems" under the current provisions. In section 3.3.5, I address how the European Commission can change the interpretive premises by altering the text of the AIA.

3.3.4.1 High-risk classification pursuant to Article 6(2) – an overview

"In addition" to the systems that are high-risk under Article 6(1), Article 6(2) decides that "AI systems referred to in Annex III shall be considered to be high-risk". Annex III recognizes AI systems in any of eight listed "areas" as high-risk. One of these areas are "critical infrastructure", which is the only alternative standing out as relevant to consider in the case of ANS deployed in commercial cargo shipping. Article 6(3) sets out exceptions to Article 6(2), establishing criteria for when "an AI system referred to in Annex III shall not be considered to be high-risk".

3.3.4.2 Annex III, Section 2, "critical infrastructure" – what are the conditions?

In the area of critical infrastructure, Section 2 of Annex III recognizes AI systems "intended to be used as safety components in the management and operation of critical digital infrastructure, road traffic, or in the supply of water, gas, heating or electricity" as high-risk.

The general concept of "safety components" is legally defined in the AIA Article 3(14). However, the Recitals also provide a definition aimed specifically at Annex III, Section 2. This recognizes that "safety components" of the designated critical infrastructures in Annex III "are systems used to directly protect the physical integrity of critical infrastructure or the health and safety of persons and property but which are not necessary in order for the system to function".^{(16) The AIA Recitals para. 55.} The use of the word "system" twice is clearly unfortunate. The Recitals must logically be interpreted to refer to the AI system (the "safety component") the first time, and the relevant critical infrastructure entity the second time. This understanding is supported in legal literature.^{(17) Paul Voigt and Nils Hullen, The EU AI Act: answers to Frequently Asked Questions (1st edition, Springer 2024) Sub-Chapter 3.1.2.2.}

Read in light of the Recitals, Annex III, Section 2 sets out three conditions that must be met for an AI system to be considered as high-risk pursuant to Article 6(2). First, the AI system must fulfill a function related to protecting the physical integrity of infrastructure. Second, the system must not be necessary for the functioning of that infrastructure. Third, the infrastructure must be a type of "critical infrastructure" within the scope of Annex III, Section 2. In the following, I will apply these conditions to ANS.

3.3.4.3 Applying the conditions of infrastructure protection and non-necessity

First, the ANS must "directly protect the physical integrity" of critical infrastructure.^{(18) The AIA Recitals para. 55.} This will undoubtedly be the case, for instance through collision and grounding avoidance.

Second, the ANS must not be "necessary" for the critical infrastructure to "function".^{(19) The AIA Recitals para. 55.} Essentially, the ANS must only add qualities to already functioning critical infrastructure entities, and these qualities must essentially be intended to ensure or enhance safety.

On the one hand, MASS depend on ANS to function once the systems are put into use during a voyage, as they are making the decisions that navigate the ship safely from point A to point B. This may indicate that, when in use, ANS could be considered "necessary".

On the other hand, one could argue that ANS, in many cases, simply represent an opportunity for shipowners to enhance the safety of their vessels and voyages. Using ANS is not "necessary" in commercial cargo shipping as such, since the conventional ships of today only use INS, or other systems for navigation. Furthermore, ANS will arguably not be "necessary" in the operation of the MASS of the future either, as these will allow for the Master to switch back to human-based navigation at any time during a voyage.^{(20) Which will certainly be required by the MASS Code, see the draft MASS Code Section 17.5.} This must be seen in connection with one of the main ideas of autonomous shipping development, which is safety enhancement and reduction of accidents. This is illustrated both by the focus areas of specific MASS projects,^{(21) T. Porathe, Å. Hoem, Ø. Rødseth, K. Fjørtoft and S. Johnsen, At least as safe as manned shipping? Autonomous shipping, safety and “human error” (SINTEF and NTNU 2018);Jiri de Vos, Robert G. Hekkenberg and Osiris A. Valdez Banda, 'The Impact of Autonomous Ships on Safety at Sea – A Statistical Analysis' (2021) 210 Ocean Eng. 107558, Section 5.} and by the broader rationale underpinning the development of the MASS Code.^{(22) The draft MASS Code Preamble para. 3.} These arguments strongly support that ANS fall within the category of safety-enhancing but non-necessary AI systems. Accordingly, the most plausible interpretation is that ANS satisfy the second condition.

3.3.4.4 Applying the third condition – is autonomous shipping covered critical infrastructure under Annex III, Section 2?

The question is whether ANS qualify as safety components of the critical infrastructures listed in Annex III, Section 2. Essentially, the question is whether autonomous ships and shipping are covered infrastructures.

Pursuant to the wording, the Annex only recognizes "critical digital infrastructure, road traffic" and "the supply of water, gas, heating or electricity" as covered critical infrastructures. In its position paper to the AIA proposal from the European Commission, Norway advocated that "AI systems for autonomous shipping" should be covered.^{(23) Government of Norway, Ministry of Local Government and Modernisation 'Norwegian Position Paper on the European Commission’s Proposal for a Regulation of the European Parliament and of the Council Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act) and Amending Certain Union Legislative Acts (COM(2021) 206)' (2021) Section 5.} Although this was not reflected in the final wording, the question remains whether one of the specifically listed categories in Annex III nonetheless encompasses autonomous shipping.

ANS are not safety components in "road traffic", nor are they likely to be considered safety components in "critical digital infrastructure". The latter concept is not defined in the AIA, but the Recitals refer to "critical digital infrastructure as listed in point (8)" of the Annex to the Critical Entities Resilience Directive (CER).^{(24) The AIA Recitals para. 55, referring to the Directive (EU) 2022/2557 (CER).} The CER refers to legal entities responsible for "digital infrastructure" in the Directive (EU) 2022/2555 (NIS2) Article 6. This may give an indication of what such infrastructure may be.^{(25) Voigt and Hullen (n 103) Sub-Chapter 3.1.2.2.}

The NIS2 Article 6 concerns physical infrastructures in digital services.^{(26) These are infrastructures such as internet exchange points, domain name systems, and cloud computing systems.} Within MASS operations, the closest analogue is likely to be ROCs. Although ROCs will depend on the physical infrastructures in NIS2 to remotely operate and monitor MASS, the ROCs will likely not qualify as such infrastructures themselves. However, even if they did, ANS would then no longer be considered their "safety components".^{(27) Systems providing fire alarm control, or similar, are more likely to qualify as safety components of the ROCs, as exemplified with cloud computing systems in the AIA Recitals para 55.}

Therefore, the decisive question is whether ANS can be considered safety components in the "supply of water, gas, heating, or electricity". The alternative of interest is gas suppliance, as it raises the question of whether the shipping of liquefied natural gas (LNG) and liquefied petroleum gas (LPG) constitute such.

In order to reduce costs and improve safety, LNG and LPG carriers of the future may be equipped with ANS. For instance, Kongsberg Maritime in Norway and Samsung Heavy Industries in South Korea are currently collaborating on a design for an autonomous LNG carrier,^{(28) Ship Technology, SHI and Kongsberg partner on autonomous LNG carrier design (21 March 2023).} and Polish researchers have performed collision avoidance tests with positive results with an autonomous LNG carrier model in quasi-real conditions.^{(29) Zbigniew Pietrzykowski, Piotr Wołejsza, Łukasz Nozdrzykowski, Piotr Borkowski, Paweł Banaś, Janusz Magaj, Jarosław Chomski, Marcin Mąka, Sylwia Mielniczuk, Anna Pańka, Paulina Hatłas-Sowińska, Eric Kulbiej and Magdalena Nozdrzykowska, 'The autonomous navigation system of a sea-going vessel' (2022) 261 Ocean Eng. 112104, Section 3.2.}

The concept of supplying gas is neither defined nor exemplified in the text of the AIA, the Recitals, or the Explanatory Memorandum. However, the AIA defines the general concept of "critical infrastructure" in Article 3(62). It does so by referring to Article 2(4) of the CER. The CER may therefore shed light on how to correctly interpret the more specific critical infrastructures of the AIA Annex III, such as gas suppliance. Legal authors are of the same impression.^{(30) Voigt and Hullen (n 103) Sub-Chapter 3.1.2.2.}

Article 2(4) of the CER generally defines "critical infrastructure" as "an asset, a facility, equipment, a network or a system", or a part of such, "which is necessary for the provision of an essential service". Through a delegated act from the European Commission in 2023, the CER generally recognizes freight transport at sea as an "essential service" under Article 2(4).^{(31) Commission Delegated Regulation (EU) 2023/2450.} Consequently, LNG and LPG shipping qualify within the general category of "critical infrastructure" in both the CER and the AIA. One may argue that considering LNG and LPG shipping as "critical infrastructure" under the specific gas suppliance alternative in the AIA Annex III will simply respect the conceptual link that the AIA Article 3(62) sets out between the AIA and the CER.

In the specific gas subsector, the Annex of the CER refers to critical infrastructures by pointing to certain definitions in Article 2 of the Directive 2009/73/EC (Gas Directive). This directive was repealed and replaced by the Directive (EU) 2024/1788 in August 2024. However, there are no decisive differences between the two Gas Directives with regard to the question at hand. Although both concern, for instance, the "supply" of LNG in the EU, this does not encompass LNG shipping. At the same time, this does not necessarily exclude LNG and LPG shipping from being considered gas suppliance under the AIA Annex III, Section 2. The positive examples of critical infrastructures in the gas subsector set out in the Gas Directives must be understood as just that – examples of what is critical infrastructure, and not necessarily a basis to conclude what it is not under the AIA.

In contrast, the Recitals of the AIA specify that "critical digital infrastructure" should be considered "as listed" in the CER Annex, leaving little room for extensive interpretation.^{(32) The AIA Recitals para. 55.} The Recitals do not refer to the CER in this way when it comes to gas suppliance. This suggests that there is room to perform a broader interpretation of the gas suppliance concept. Such an interpretation, however, must arguably align with the objectives of the AIA.

The AIA is intended to ensure a "high level of protection" for health and safety.^{(33) The AIA Article 1(1).} Consistent with this, the Recitals focus on how the significance of critical infrastructure becomes evident in society if the safety components of the infrastructure fail or malfunction. According to the Recitals, such failure or malfunctioning may lead to "appreciable disruptions in the ordinary conduct of social and economic activities".^{(34) The AIA Recitals para. 55.} More importantly, it can result in "serious harm to the provision of basic supplies to the population", which may lead to "an imminent threat to life or the physical safety of a person".^{(35) The AIA Recitals para. 33.}

Several European countries import large amounts of gas for heating and production of electricity. The gas is most commonly transported through pipelines, but in some supply chains, it is also transported as LNG and LPG at sea.^{(36) The International Energy Agency, Gas imports and exports (2025).} The failure of ANS in LNG and LPG shipping can lead to accidents with severe damage to the MASS, crew, and other ships, in addition to serious environmental damage. Furthermore, malfunctions in ANS could affect and delay multiple LNG and LPG carriers simultaneously, which may prevent significant amounts of gas from reaching its destinations on time.^{(37) Such simultaneous failure may, for instance, be due to many ships, with different shipowners, being connected to the same autonomous navigation system or ROC, where there is an error, or is subject to a successful cyberattack. See Section 2.4 of the thesis regarding designs of ANS.} Such delays may have immediate disruptive effects on European societies. Illustrated by the European challenges following the Russian invasion of Ukraine, the International Energy Agency (IEA) emphasized that "reliance on imported natural gas for electricity can lead to painful price spikes when supplies are disrupted".^{(38) The International Energy Agency, Natural gas in electricity generation (2025).}

In conclusion, an interpretation of Annex III, Section 2, suggests that ANS may be regarded as safety components in critical infrastructure when that infrastructure is LNG or LPG shipping. In such a scenario, the ANS would qualify as high-risk AI systems under Article 6(2).

3.3.4.5 The exception to Article 6(2) – can the European Commission rule out ANS as high-risk AI systems under Article 6(3)?

The chapeau of the AIA Article 6(3) decides that when an AI system "does 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", it shall not be considered high-risk under Article 6(2). The provision further identifies four situations in which this condition is presumed to be fulfilled. As it is summarized in legal literature, the systems in the four listed cases should only support human decision or activity.^{(39) Voigt and Hullen (n 103) Sub-Chapter 3.1.2.10.} In such cases, de facto decision-making power remains with humans, while the AI systems merely assist the humans as a tool.

For the purpose of this thesis, ANS are capable of making navigational decisions independently, without necessarily requiring approval from a human being.^{(40) Thesis Section 2.1.2 and the Draft MASS Code Section 4.7.} Accordingly, ANS are unlikely to qualify for an exception from Article 6(2) pursuant to Article 6(3).