Swiss AI Providers and the Territorial Reach of the AI Act

Die KI-Verordnung der EU erfasst Anbieter und Betreiber aus Drittstaaten, wenn die von ihren KI-Systemen hervorgebrachte Ausgabe in der Union verwendet wird. Für Schweizer Anbieter stellt sich damit eine besonders heikle Frage der Territorialität. Der Beitrag untersucht Art. 2 Abs. 1 Bst. c KI-VO im Lichte der Entstehungsgeschichte, unterscheidet die technische Verwendung einer Ausgabe von ihrer rechtlichen Wirkung und plädiert für eine Auslegung, die auf Vorhersehbarkeit und einem hinreichend engen Bezug zur Union beruht.

I. Introduction

A Geneva-based company supplies a French bank with a scoring tool designed to assess the creditworthiness of loan applicants. The model is developed in Switzerland, the IT infrastructure may be located in several countries, and the provider has no establishment in the Union. Yet the credit decision is taken in Lyon on the basis of the score produced by the system. Can Regulation (EU) 2024/1689 on artificial intelligence (AI Act) apply to the Swiss provider? Article 2(1)(c) answers in principle in the affirmative where the ‘output’ produced by the AI system is used in the Union.[1]

This formula is one of the most ambitious territorial choices of the AI Act. It depends neither on the provider’s seat, nor on the place of development, nor on the location of the servers. It shifts attention to the operational result of the system. For Switzerland, which is a member neither of the Union nor of the European Economic Area, the question is therefore not whether the AI Act will be incorporated into Swiss law, but whether it applies to Swiss operators by virtue of a functional connection to the Union market or to uses located in the Union.[2]

The issue has become very concrete. Swiss companies may be subject to the AI Act when they place AI systems on the Union market or when the outputs of those systems are used in the Union; the legal analysis prepared for the Federal Council examines precisely the consequences of the AI Act for Switzerland. At the same time, Switzerland still has no general cross-sectoral AI legislation. On 12 February 2025, the Federal Council decided to ratify the Council of Europe Framework Convention on Artificial Intelligence and to implement it through targeted legislative amendments complemented by non-binding measures. A draft for consultation is to be prepared by the end of 2026, notably on transparency, data protection, non-discrimination and supervision.[3]

This contribution does not aim to provide an overview of the substantive obligations of the AI Act. It focuses on a narrower question: how should the territorial connecting factor based on the use of an output in the Union be interpreted where the provider is established in Switzerland? The difficulty is heightened by a tension within the final text: Article 2(1)(c) refers to output that ‘is used’ in the Union, whereas recital 22 refers to output that ‘is intended to be used’ in the Union.[4] The hypothesis defended here is that a purely factual reading of the word ‘used’ would be too broad. What is required is a sufficiently foreseeable link between the provider’s activity, the intended purpose of the system or its outputs, and the legal relationship in which those outputs become operative. This reading preserves the effectiveness of the AI Act without turning every global circulation of an output into an automatic connection to the Union.

II. Swiss Providers within the Territorial Scope of the AI Act

The starting point is Article 2(1) AI Act. Three connecting factors are particularly important. First, the Regulation applies to providers placing on the market or putting into service AI systems in the Union, irrespective of where they are established. Second, it applies to deployers established or located in the Union. Third, it covers providers and deployers established or located in a third country where the output produced by the system is used in the Union. These three scenarios do not overlap: the first rests on market access, the second on the location of the professional user, the third on the destination of the output.[5]

For a Swiss provider, Article 2(1)(a) may suffice as soon as a system is marketed in the Union. The obligations relating to high-risk systems then include, for providers established in third countries, the appointment of an authorised representative in the Union.[6] Likewise, providers of general-purpose AI models established in third countries must, in the cases provided for by the Regulation, appoint an authorised representative before placing the model on the Union market.[7] The logic is classical: access to the internal market justifies the application of European rules and organises an institutional relay for their enforcement. Since the Digital Omnibus on AI, the date from which the obligations for high-risk systems apply has, however, been deferred.[8]

Point (c) is different. It may apply even where the Swiss provider has not itself placed the system on the Union market. This is precisely what makes the provision difficult. A system may be developed for a US company, integrated into a global service and produce an output that is ultimately used by a European subsidiary. The provision does not expressly specify whether such use must have been intended, foreseen or merely possible. Nor does it define the intensity of the link required between the third-country provider and the European use.

This openness explains why the extraterritoriality of the AI Act is already the subject of a critical literature. An empirical study published in 2026 starts from the premise that the AI Act applies to any AI system whose output is used in the Union and finds that companies without an EU establishment refer to its prohibitions far less often than companies established in the Union – a gap the authors attribute to limited extraterritorial enforcement and to uncertainty about the Act’s scope.[9] The question is thus not only one of the Union’s prescriptive jurisdiction, but also of the foreseeability of the connecting factor for economic operators.

The architecture of the Regulation also reveals a gap between territorial scope and the mechanisms ensuring a presence in the Union. Article 22 requires a third-country provider to appoint an authorised representative before making a high-risk AI system available on the Union market. That obligation, however, is tied to making the system available on the market. Where Article 2(1)(c) applies solely because an output is used in the Union, without the provider concerned placing anything on the market, the Regulation establishes no equivalent general mechanism of representation. This asymmetry confirms that point (c) targets situations further removed from the classical market-access pattern and makes it all the more necessary to define the required territorial link precisely.[10]

III. From ‘Used’ to ‘Intended to Be Used’: A Revealing Legislative History

The choice of words is not a drafting detail. The Commission’s 2021 proposal already provided that the Regulation would apply to providers and users located in a third country where the output produced by the system is ‘used in the Union’.[11] The Council’s general approach of December 2022 retained this logic and the same wording.[12] The institutional starting point was therefore a criterion based on the actual use of the output, without any express condition relating to its anticipated destination.

The European Parliament, however, proposed a substantial shift. In its amendments of 14 June 2023, it replaced, in Article 2(1)(c), the words ‘is used in the Union’ with ‘is intended to be used in the Union’, and made the same change in the corresponding recital.[13] This version introduced a prospective dimension: the connection no longer depended solely on whether an output ultimately crossed the border, but on its envisaged destination.

The final text presents a peculiarity. The operative provision of Article 2(1)(c) returned to ‘is used in the Union’. Yet recital 22 of the adopted Regulation retains the expression ‘is intended to be used in the Union’, in a sentence explaining that the extension to third-country providers and deployers serves to prevent the circumvention of the Regulation and to ensure effective protection of natural persons located in the Union.[14] The coexistence of these two formulations precludes reading one too quickly in the light of the other: a recital cannot rewrite the provision, but it sheds light on the anti-circumvention purpose of the rule and on the nature of the territorial link the legislature intended to capture.

This tension supports an intermediate interpretation. Article 2(1)(c) does not require proof of a subjective intention on the part of the provider to target the Union: the word ‘intended’ does not appear in the operative rule. But recital 22 makes it difficult to accept an interpretation under which any incidental, unforeseeable or wholly extraneous use would suffice. The intended purpose of the system, contractual stipulations, categories of customers, the commercial territory targeted and reasonable knowledge of uses thus become relevant indicators in determining whether the European use genuinely falls within the connection the Regulation was designed to capture.

For a Swiss provider, this nuance is decisive. A Zurich company should not be subject to the AI Act merely because a non-European client subsequently exports a report to the Union without that use having been foreseen or being reasonably foreseeable. The situation is different where the product is designed for European activity, where the contract authorises or organises such use, where data or parameters specific to the Union market are built in, or where the provider directly supports deployment in a Member State. The foreseeability criterion therefore does not neutralise extraterritoriality; it helps to distinguish a functional connection from an accidental presence.

IV. From Technical Location to Legal Effect

The choice of ‘output’ as a connecting factor is all the more significant because the notion lies at the heart of the very definition of an AI system. Article 3(1) describes such a system as generating outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments.[15] The category is thus functional and deliberately broad: an output is not merely an automated decision in the strict sense, but may be a recommendation, a ranking, a score or content subsequently incorporated into a human decision.

This breadth explains why the place of use cannot be identified uniformly. A prediction may be consulted in several States; a recommendation may be incorporated into a decision-making process elsewhere; content may be generated in Switzerland and disseminated simultaneously throughout the Union. Territoriality cannot then rest on mere data transmission. What matters is the function the output performs in the regulated activity or in the legal relationship concerned.

The vocabulary of the AI Act itself thus invites a distinction between generation, transmission and use. Generation describes the operation of the system; transmission describes the circulation of the result; use presupposes that an actor relies on that result in an activity. For Article 2(1)(c), it is this third stage that establishes the connection. In a Swiss–EU context, it prevents the location of the model or the servers from obscuring the real centre of the operation, while leaving open the question of when that use is sufficiently attributable to the provider established outside the Union.

Cross-border algorithmic systems make traditional geographical criteria less intuitive. The place where the model is developed, where it is hosted, where data are processed and where a decision is taken may all differ. A service may change its cloud infrastructure without any change in the economic or legal relationship the system serves. In these circumstances, tying the territorial scope of the AI Act to an element of technical infrastructure would produce an unstable and easily manipulable territoriality.

Article 2(1)(c) breaks precisely with this logic. It does not ask where the algorithm ‘is’, but where its output is used. This shift deserves to be taken seriously. It brings regulatory territoriality closer to the function of the system in a concrete relationship: a credit score matters when it enters into a financing decision; a ranking of applications when it influences a recruitment procedure; a clinical recommendation when it is relied on in a patient’s care.

Not every consultation of an output in the Union should suffice, however. A distinction is needed between the technical use of an output and its legal or economic effect. Suppose a Swiss provider develops a model for a Canadian client, and an executive travelling in Paris occasionally consults the dashboard. A literal reading might consider that an output has been ‘used’ in the Union. It would nevertheless be difficult to infer that the Swiss provider ought to have organised its entire compliance ex ante around EU law. Conversely, where a system is designed to assess credit applications from French consumers, the connection with the Union is structural and foreseeable, even if the model is hosted in Zurich.

The idea of a legally operative effect makes it possible to order these situations. It does not replace the text of the AI Act and does not create a new territorial criterion. It helps to determine whether the European use of the output corresponds to the economic and legal purpose of the system or is merely an accidental fact. The decisive point then becomes the relationship in which the output is meant to be taken into account.

This approach is not foreign to EU digital law. The GDPR already established a territoriality that goes beyond the place of establishment, notably where persons in the Union are targeted or their behaviour in the Union is monitored. The guidelines of the European Data Protection Board insist on the existence of a targeting element and give precisely the example of a Swiss university whose online activity may, depending on the circumstances, remain outside the scope of the GDPR or fall within it if it specifically targets persons in the Union.[16] The parallel should not be pushed too far, since Article 2 AI Act uses different wording. It nevertheless shows that European extraterritoriality is not necessarily conceived as a consequence of the mere global accessibility of a service.

V. Foreseeability as an Interpretive Limit

Foreseeability plays a dual role. It protects operators against a purely fortuitous connection, and it strengthens the legitimacy of an extraterritorial rule. A Swiss provider that sells a recruitment tool to a German group, or designs a scoring system for French institutions, can reasonably anticipate that its outputs will be used in the Union. By contrast, a provider whose service is intended for a third-country market and which does not control a marginal subsequent use in the Union is in a different position.

For Article 2(1)(c), the AI Act contains no general ‘substantial connection’ clause comparable to that of the Digital Services Act (DSA). The DSA applies to providers offering services in the Union, which presupposes a substantial connection to the Union; the legislature refers in particular to a significant number of recipients in one or more Member States or to the targeting of activities towards them.[17] This textual difference precludes any mechanical transposition of the DSA criteria to the AI Act. It nevertheless provides a useful point of comparison: when the Union extends its rules to actors located abroad, the density of the link with the internal market remains a structural concern.

In the context of the AI Act, that density may be assessed through several indicators: the contractual purpose of the system; the sectors and territories for which it was trained or configured; the languages and regulatory parameters built into it; the existence of customers or group entities in the Union; commercial documentation; the provider’s ability to foresee or control the use of the output; and, finally, the legal relationship the output is meant to influence. None of these indicators should be decisive on its own.

This framework avoids two excesses. The first would be to render Article 2(1)(c) practically inoperative by requiring direct marketing in the Union, whereas the provision specifically targets situations in which the output crosses the border without the system itself necessarily being placed on the market there. The second would be to treat any material use in the Union as sufficient, even where it is unforeseeable for the provider. Between these two poles, foreseeability and a sufficiently close connection offer interpretive criteria consistent with the objective of the text.

VI. Three Swiss–EU Configurations

1. Credit. A Geneva company develops a scoring engine for a French bank. The data are processed in Switzerland and the score is transmitted to the bank’s office in Lyon, which uses it to accept, refuse or price a loan. Here, the application of the AI Act should hardly be in doubt if the Regulation’s substantive conditions are otherwise met. The output is not only used in the Union; its normal function is precisely to contribute there to a decision concerning a customer. The Swiss location of the provider or of the infrastructure does not neutralise that connection.

2. Recruitment. A Zurich company supplies a US multinational with an application-screening tool. The main contract is concluded in the United States, but the system is then used by a German subsidiary to recruit in Munich. The situation is more delicate if the Swiss provider neither negotiated nor configured the European use. The analysis should focus on the structure of the service: a worldwide licence, the announced purpose, knowledge of the subsidiaries using the tool, linguistic and territorial configuration, and support provided to the subsidiary. The more these elements make the German use identifiable, the more foreseeable the application of Article 2(1)(c) becomes.

3. Health. A Basel developer supplies a diagnostic support tool to a Swiss hospital network. A physician of that network remotely treats a patient located in Italy and relies on the system’s recommendation in the patient’s care. The mere presence of the patient in the Union should not automatically settle the question. A distinction must be drawn between a situation in which cross-border services form part of the network’s organised activity and an isolated case. Here again, the use of the output must be placed within the legal and professional relationship that gives it its significance.

These examples show that the notion of output ‘used in the Union’ cannot be reduced to the place where a screen is consulted. It more usefully refers to the place where the result enters a decision-making or operational process within an identifiable activity. This reading remains faithful to the legislature’s choice of the output as a connecting factor, while avoiding an unlimited extension based on accidental uses.

VII. The Swiss Defendant: Brussels I bis or the Lugano Convention?

The case of a provider established in Switzerland calls for a clarification that goes beyond the territorial scope of the AI Act. The application of the EU Regulation to a Swiss operator does not mean that the Brussels I bis Regulation automatically governs jurisdiction over it. Under Article 6(1) Brussels I bis, where the defendant is not domiciled in a Member State, the jurisdiction of the courts of each Member State is in principle determined by the law of that State, subject to certain protective or exclusive heads of jurisdiction.[18] For a provider domiciled in Switzerland, however, account must be taken of the 2007 Lugano Convention, which binds Switzerland and the European Union and is the treaty-based counterpart of the Brussels I system.[19]

This distinction is particularly important for the analysis of cross-border algorithmic decisions. Article 5(3) of the Lugano Convention provides that a person domiciled in a State bound by the Convention may be sued, in matters relating to tort, delict or quasi-delict, in the courts for the place where the harmful event occurred or may occur.[20] The localisation problem therefore does not disappear when the provider is Swiss: it moves from the Brussels I bis Regulation to a parallel instrument whose logic remains close. Under Protocol 2 to the Convention, courts applying it must pay due account to the relevant case law of the Court of Justice, including on the place where the harmful event occurred.[21] The interpretation of that notion thus remains fully relevant to determining whether a provider based in Geneva, Zurich or Basel may be sued before a court of a Member State on account of an output used in the Union.

This reinforces the relevance of the notion of a legally operative effect outlined in section IV. If a system designed in Switzerland produces a recommendation that is incorporated into a credit decision in France, a recruitment decision in Germany or a care decision in Italy, jurisdiction should not depend on the server or on the place of computation. In a non-contractual action, the decisive question is rather where the direct damage materialised within the legal relationship affected. The Lugano Convention thus reproduces, for Swiss–EU litigation, a problem very close to the one raised by Article 7(2) Brussels I bis within the Union.

The position is different as regards the applicable law. The Rome II Regulation has universal application: the law it designates applies even if it is the law of a third State.[22] A French or German court seised of a non-contractual action against a Swiss provider may therefore apply Rome II to determine the applicable law, without the defendant’s Swiss domicile standing in the way. Article 4(1) designates in principle the law of the country in which the damage occurs, irrespective of the country in which the event giving rise to the damage occurred and of the countries in which indirect consequences occur.[23] A single dispute may thus involve three distinct layers: the AI Act for the regulatory scope, the Lugano Convention for jurisdiction, and Rome II for the applicable law.

This layering shows why regulatory extraterritoriality and private international law should neither be confused nor studied in isolation. Article 2(1)(c) AI Act may subject a Swiss provider to European obligations because its output is used in the Union; the Lugano Convention then determines whether that provider may be sued before a court of a State bound by it; Rome II may finally lead to the application of the law of the State in which the direct damage occurs. The same territorial fact – the legally operative use of an output – may therefore play different roles depending on the instrument, without thereby becoming a single autonomous criterion.

VIII. The AI Act Does Not Replace Private International Law

The extraterritoriality of the AI Act must not be confused with rules on jurisdiction or applicable law. Saying that a Swiss provider falls within the territorial scope of the Regulation does not, in itself, mean that a particular court has jurisdiction over a civil action, or that the law of a Member State governs the relationship as a whole. This distinction is particularly important for private enforcement.

Marion Ho-Dac pointed out early on that the AI Act, despite its architecture of regulatory and product-safety law, raises questions of private international law as soon as it operates within transnational value chains and may give rise to claims between private parties.[24] The classical instruments retain their function here. For disputes in matters relating to tort, Article 7(2) Brussels I bis and the Rome II Regulation may become decisive; for contractual relationships, Rome I retains its role. The territorial scope of the AI Act and the connecting factors of private international law answer different questions.

This dissociation is particularly visible where the provider is Swiss. The AI Act may impose compliance obligations because an output is used in the Union, while a civil action against the provider remains subject to the ordinary rules on jurisdiction and conflict of laws. Conversely, a court of a Member State may have jurisdiction over a dispute relating to an algorithmic decision without all the obligations of the AI Act necessarily applying to the operator concerned. Courts should therefore avoid turning the territorial scope of an economic regulation into a general rule of international jurisdiction.

The European case law on the localisation of damage offers methodological guidance here. In several fields, the Court of Justice distinguishes the event giving rise to the damage from the place where the initial damage materialises, and refuses to take into account mere subsequent repercussions.[25] This case law obviously does not determine the meaning of Article 2 AI Act. It nevertheless shows the caution with which EU law treats effects-based connecting factors. The relevant effect must be legally qualified, not merely observed.

IX. The Swiss Specificity: Regulatory Autonomy and Exposure to EU Law

Switzerland offers particularly fertile ground for this discussion. It remains outside the legal order of the Union while being closely integrated into its economic environment. In the field of AI, the Federal Council has chosen not simply to adopt the European horizontal model. The federal approach combines targeted legislative amendments with non-binding measures, notably in the context of implementing the Council of Europe Framework Convention.[26]

Switzerland signed the Convention on 27 March 2025. It is the first legally binding international treaty devoted to AI, human rights, democracy and the rule of law. Its orientation differs from that of the AI Act: it lays down a framework of safeguards and State obligations, whereas the EU Regulation establishes a detailed regime applicable to operators and categories of systems. For Swiss companies, however, the two developments may meet in practice.[27]

This situation reinforces the value of a precise reading of Article 2(1)(c). Switzerland cannot be treated as a mere periphery of EU law. But its regulatory autonomy does not prevent Swiss companies from being exposed to European norms when they participate in value chains oriented towards the Union. The challenge is therefore less to choose between autonomy and alignment than to identify, transaction by transaction, the legally relevant points of contact.

For companies, this approach calls for a mapping of uses rather than a mapping of servers. They need to know where their systems are offered, who deploys them, in which States the outputs are intended to be used, which decisions they feed and which group entities are involved. Clear contractual governance can also reduce uncertainty: definition of the territory of use, information obligations in the event of redeployment, geographical restrictions, documentation of purposes and cooperation with professional users.

X. Towards a Criterion of a Sufficiently Close and Foreseeable Connection

Article 2(1)(c) is briefly worded, but it is not condemned to a maximalist interpretation. Its function is to prevent an operator from escaping the AI Act by artificially keeping its entire technical organisation outside the Union while its outputs are intended to be used there. That purpose is legitimate. It does not require every output that fortuitously crosses a digital border to be brought within the Regulation.

A balanced interpretation could rest on three converging elements. First, an identifiable use of the output in the Union, and not mere accessibility. Second, a functional link between that use and the purpose of the system or the economic relationship in which it is supplied. Third, a reasonable foreseeability of that use for the third-country provider or deployer. The aim is not to turn these indicators into autonomous conditions that the legislature did not write. They serve to give a legally operational content to the word ‘used’ in the light of recital 22, of the anti-circumvention objective and of the requirement of legal certainty.

This method brings the territoriality of the AI Act and the logic of private international law closer together without merging them. In both fields, mere physical geography no longer suffices when activities are distributed. The law looks for connecting factors capable of expressing a sufficiently real relationship between a situation and a legal order. In the AI Act, the choice of the output as a connecting factor already signals that the legislature favours functional effect over the location of infrastructure.

The literature on the ‘Brussels effect’ also invites a distinction between extraterritorial legal application, voluntary corporate adoption of European standards, legislative emulation by third countries and the incorporation of EU standards into international agreements and technical standards.[28] In Switzerland, these phenomena coexist. Some companies will be directly subject to the AI Act; others will align their products to preserve smooth access to the European market; and the Swiss legislature will develop its own framework against the background of an already dense European regulatory environment.

XI. Foreseeability in Practice: Contracts, Documentation and Redeployment

Foreseeability is not only an interpretive tool for courts. It can also be documented ex ante by operators. For a Swiss provider, licence or service agreements should identify the envisaged territories of use, the categories of deployers, the authorised uses and the conditions for any redeployment to the Union. Such documentation does not by itself determine the scope of the AI Act, which remains a matter of mandatory law, but it makes it possible to establish what the provider could reasonably foresee when the system was made available.

Contracts therefore cannot serve as an avoidance mechanism. A clause formally prohibiting any use in the Union should not suffice if, in practice, the provider configures the system for European customers, processes data specific to their activities or provides support directly linked to deployment in the Union. Conversely, where European redeployment occurs without the provider’s knowledge and in breach of geographical restrictions that are actually enforced, the situation calls for a different assessment. Technical documentation, access logs, territorial settings and exchanges with the customer may then become evidence of whether the European use was foreseeable.

This approach is particularly suited to AI value chains, in which an actor’s role may change after the initial delivery. A system supplied in Switzerland may be integrated into a broader service, resold, adapted or used by a group company in the Union. Territorial compliance must therefore be monitored over time. For the provider, the challenge is less to freeze a geographical map than to identify the changes in use that may bring the system within the scope of the AI Act.

Such governance reflects the logic defended in this contribution: the relevant connection results from the combination of a concrete use of the output, its function in an identifiable relationship and the reasonable possibility for the operator to attribute that use to the Union. It also gives the proposed criterion an operational dimension, without turning it into an autonomous condition added to the European text.

XII. Conclusion

For Swiss providers, the territorial question raised by the AI Act cannot be reduced to the alternative ‘inside the Union / outside the Union’. Article 2(1)(c) adopts a more functional connecting factor: it looks at the use of the output. This solution is well suited to distributed systems, but it creates an area of uncertainty where the European use is indirect, occasional or difficult for the provider to control.

The interpretation proposed here distinguishes the mere technical presence of an output from its insertion into an identifiable relationship in which it becomes legally or economically operative. The foreseeability of that use and the existence of a sufficiently close connection with the Union should guide this analysis. Such a reading does not add to the text a condition foreign to it; it gives the notion of ‘use’ a scope compatible with legal certainty and with the function of the AI Act.

Switzerland makes this tension particularly visible. Its regulatory autonomy remains intact, but its AI providers are embedded in cross-border chains whose effects unfold daily in the Union. The question is therefore no longer where the algorithm is located. It is where, for what purpose and with what degree of foreseeability its output enters a legal relationship that EU law seeks to protect.

The legislative history finally confirms that this question is not artificial. Between the Commission’s proposal, the Council’s position, the Parliament’s amendment and the final compromise, the legislature hesitated between the actual use and the intended destination of the output. The adopted text retains both ideas, one in the Article and the other in the recital. For Swiss operators, this tension calls for a reading that protects the market and persons in the Union without detaching territoriality from all economic and legal foreseeability.

Volodymyr Ponomarov holds a Master’s degree in International and European Law from the University of Geneva. He works as a legal assistant at Bergwill Group in Geneva and is a doctoral candidate in private international law at Jean Moulin Lyon 3 University. His research focuses on the territoriality of algorithmic systems, European private international law and private enforcement. Contact: vlad@ponomarov.ch.

References

[1] Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), OJ L, 2024/1689, 12 July 2024, Art. 2(1)(c), as amended by Regulation (EU) 2026/1744 (consolidated version of 27 July 2026).

[2] On the distinction between the extraterritorial application of the AI Act and other forms of EU regulatory influence on third countries, see Graham Greenleaf, EU AI Act: Brussels Effect(s) or a Race to the Bottom?, (2024) 190 Privacy Laws & Business International Report 1, 3–6.

[3] Federal Office of Communications (OFCOM), Intelligence artificielle, https://www.bakom.admin.ch/fr/intelligence-artificielle (accessed 6 October 2026), referring to the Federal Council’s decision of principle of 12 February 2025 and to the legal analysis of the Council of Europe Convention and of the AI Act and their consequences for Switzerland. According to the same source, the draft for consultation is to be prepared by the Federal Department of Justice and Police in cooperation with DETEC and the FDFA.

[4] AI Act (n. 1), Art. 2(1)(c) and recital 22.

[5] AI Act (n. 1), Art. 2(1)(a)–(c).

[6] AI Act (n. 1), Art. 22(1).

[7] AI Act (n. 1), Art. 54(1).

[8] Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 as regards the simplification of the implementation of harmonised rules on artificial intelligence (Digital Omnibus on AI), published in the Official Journal on 24 July 2026. The obligations for high-risk AI systems under Art. 6(2) and Annex III AI Act now apply from 2 December 2027, and those for high-risk systems under Art. 6(1) and Annex I from 2 August 2028.

[9] Kamil Szostak/Gijs van Dijck/Konrad Kollnig, Mind the Competitiveness Gap: Measuring the AI Act’s Extraterritorial Reach, Computer Law & Security Review 62 (2026), 106357, DOI 10.1016/j.clsr.2026.106357.

[10] AI Act (n. 1), Art. 22(1); for general-purpose AI models placed on the Union market, see also Art. 54.

[11] European Commission, Proposal for a Regulation of the European Parliament and of the Council laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), COM(2021) 206 final, 21 April 2021, Art. 2(1)(c).

[12] Council of the European Union, General approach on the Artificial Intelligence Act, doc. 14954/22, 25 November 2022 (adopted on 6 December 2022), Art. 2(1)(c).

[13] Amendments adopted by the European Parliament on 14 June 2023 on the proposal for a regulation laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), P9_TA(2023)0236, OJ C, C/2024/506, 23 January 2024, amendments to Art. 2(1)(c) and to recital 11 of the proposal.

[14] AI Act (n. 1), Art. 2(1)(c) and recital 22.

[15] AI Act (n. 1), Art. 3(1).

[16] European Data Protection Board, Guidelines 3/2018 on the territorial scope of the GDPR (Article 3), version 2.1, 12 November 2019, in particular the example of a Swiss university and the targeting of persons in the Union.

[17] Regulation (EU) 2022/2065 of the European Parliament and of the Council of 19 October 2022 on a Single Market for Digital Services (Digital Services Act), OJ L 277, 27 October 2022, p. 1, recitals 7–8 and Art. 3(d)–(e).

[18] Regulation (EU) No 1215/2012 of the European Parliament and of the Council of 12 December 2012 on jurisdiction and the recognition and enforcement of judgments in civil and commercial matters (Brussels I bis), OJ L 351, 20 December 2012, p. 1, Art. 6(1).

[19] Convention on jurisdiction and the recognition and enforcement of judgments in civil and commercial matters, done at Lugano on 30 October 2007, OJ L 339, 21 December 2007, p. 3 (SR 0.275.12).

[20] Lugano Convention (n. 19), Art. 5(3).

[21] Lugano Convention (n. 19), Protocol 2 on the uniform interpretation of the Convention, Art. 1(1). On the interpretation of the place where the harmful event occurred, see in particular CJEU, 5 July 2018, flyLAL-Lithuanian Airlines, C-27/17, ECLI:EU:C:2018:533, para. 27.

[22] Regulation (EC) No 864/2007 of the European Parliament and of the Council of 11 July 2007 on the law applicable to non-contractual obligations (Rome II), OJ L 199, 31 July 2007, p. 40, Art. 3.

[23] Rome II (n. 22), Art. 4(1).

[24] Marion Ho-Dac, The EU AI Act and Private International Law: A First Look, EAPIL Blog, 21 October 2024, https://eapil.org/2024/10/21/the-eu-ai-act-and-private-international-law-a-first-look/ (accessed 6 October 2026).

[25] See in particular ECJ, 19 September 1995, Marinari, C-364/93, ECLI:EU:C:1995:289; CJEU, 10 December 2015, Lazar, C-350/14, ECLI:EU:C:2015:802; on the requirement of a particularly close connecting factor in matters relating to tort, flyLAL-Lithuanian Airlines (n. 21), para. 27.

[26] Federal Council, decision of principle of 12 February 2025 on the regulation of artificial intelligence; OFCOM (n. 3).

[27] Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law, CETS No. 225, signed by Switzerland on 27 March 2025.

[28] Greenleaf (n. 2), distinguishing extraterritorial application, de facto corporate adoption, legislative emulation by third countries and adoption in international agreements and standards.