Why the EU AI Act Matters for International Arbitrations in Singapore
This article examines who the AI Act actually regulates in an arbitral context, how its reach may extend to Singapore-seated proceedings, what obligations may apply to participants outside the EU and what tribunals can do to reduce procedural and enforcement risks associated with AI use.
Regulation (EU) 2024/1689 (the “AI Act”) may matter to arbitral proceedings in two ways that do not depend solely on where the tribunal sits. First, its extraterritorial scope follows not the location of the arbitration but the connecting factors set out in Article 2, including where an AI system is placed on the EU market or put into service, where the relevant deployer is established or located and, in certain circumstances, where the system’s output is used. The use of an AI system’s output within the EU may therefore bring a particular third-country provider or deployer within the Act’s scope. Second, and more immediately consequential for practitioners, non-compliant or undisclosed AI use may undermine an award through conventional channels, including procedural-irregularity and public-policy objections under the New York Convention, the Model Law and Singapore’s International Arbitration Act. A breach of the AI Act may not itself be sufficient, but it may become relevant where the manner in which AI was used engages one of those recognised grounds.
The EU AI Act: A Risk-Based Framework Introduced in Stages
The EU AI Act entered into force on 1 August 2024. It establishes a risk-based framework under which different requirements apply depending on the type and intended purpose of an AI system. It prohibits certain practices, imposes requirements on high-risk systems, regulates providers of general-purpose AI models and introduces transparency obligations for certain systems and AI-generated content.
The Act applies in stages. Its general provisions, prohibitions and AI-literacy requirement took effect on 2 February 2025, followed by the rules for providers of general-purpose AI models on 2 August 2025. 1AI Act, Arts 4, 5 and 113; European Commission, Timeline for the Implementation of the EU AI Act. Most remaining provisions, including the transparency obligations in Article 50, became applicable on 2 August 2026. Providers of certain systems generating synthetic content that were already on the market have until 2 December 2026 to comply with Article 50(2).2AI Act, Art 50(2); European Commission, Timeline for the Implementation of the EU AI Act, above n 1.
The obligations already applicable include Article 4, which requires providers and deployers to take proportionate measures supporting AI literacy, and the targeted transparency requirements in Article 50. Article 50 covers matters such as direct interaction with AI, synthetic content, emotion-recognition systems and deepfakes; it does not impose a general duty to disclose the use of AI for legal research, translation or drafting. Certain infringements, including breaches of Article 50, may attract substantial financial penalties.3AI Act, Arts 4, 50 and 99(4)(g) and (6); European Commission, Guidelines on Transparency Obligations for Providers and Deployers of Certain AI Systems, 20 July 2026.
The timetable for high-risk systems was subsequently amended by Regulation (EU) 2026/1744. The requirements for standalone systems listed in Annex III, originally due to apply from 2 August 2026, have been postponed until 2 December 2027. Those for systems embedded in regulated products covered by Annex I will apply from 2 August 2028. 4Regulation (EU) 2026/1744, Article 1(40)(b)(i), amending Article 113 of Regulation (EU) 2024/1689; OJ L, 2026/1744, 24 July 2026.
Once the high-risk regime applies, providers will face requirements concerning risk management, documentation, data governance and human oversight, while deployers must use systems in accordance with their instructions, provide appropriate oversight, monitor their operation and retain relevant logs. Which obligations fall on an institution, tribunal or individual arbitrator will depend on whether it acts as the provider or deployer of the system.5AI Act, Arts 8–16 and 26, particularly Arts 14 and 26(6); Regulation (EU) 2026/1744, Art 1(40)(b)(i).
The Commission attributes the postponement to delays in the availability of standards and other support tools needed to implement the high-risk regime.6European Commission, AI Act Service Desk, Frequently Asked Questions – Digital Omnibus: “What are the new timelines proposed?”, stating that the revised timetable acknowledges delays in standards and other implementation support tools. Institutions, tribunals and technology providers should nevertheless use this period to identify potentially affected systems and determine how responsibility for compliance will be allocated.
Who Is the Regulated Party in an Arbitration under the AI Act?
The AI Act allocates obligations according to the role performed in relation to a particular AI system. Article 3(3) defines a “provider” as a person or entity that develops an AI system or general-purpose AI model, or has one developed, and places it on the market or puts the system into service under its own name or trademark. Under Article 3(4), a “deployer” is a person or entity “using an AI system under its authority”, except for personal, non-professional activities.7AI Act, Art 3(3) and (4). In broad terms, the provider supplies the system, while the deployer uses it professionally.
These concepts do not map neatly onto arbitration. A technology company that develops and supplies an AI tool will ordinarily be its provider. Depending on who selects, controls and operates the system, the deployer could be the arbitral institution, the tribunal, an individual arbitrator, or a party or law firm using it for its own work.
The distinction between the institution and the tribunal is important. Annex III, point 8(a), identifies as potentially high-risk certain systems intended to assist with researching and interpreting facts and law or applying the law to a concrete set of facts, including systems intended “to be used in a similar way in alternative dispute resolution”.8AI Act, Annex III, point 8(a). The Commission’s draft guidelines include commercial arbitration institutions within their description of the “ADR bodies” to which this provision may apply.9AI Act, Annex III, point 8(a); European Commission, Draft Guidelines on the Classification of High-Risk AI Systems, 19 May 2026 (“Draft Guidelines”), paras 428–429. However, 16 arbitration and dispute-resolution organisations have questioned this approach, arguing that institutions generally administer proceedings while tribunals decide the merits, and that the proposed interpretation risks attaching obligations to the wrong actor.10Joint Submission of International Arbitration Institutions in Response to the Targeted Consultation on the Draft Commission Guidelines, 23 July 2026; LCIA, LCIA Leads Cross-Institutional Response to Draft EU AI Act Guidelines, 6 August 2026. Classification should depend on the function performed. Administrative uses, such as scheduling hearings or organising documents, differ from AI used to analyse evidence, research law or assist with decisions. In the latter cases, the relevant deployer is more likely to be the institution, tribunal or individual arbitrator exercising that function. In ad hoc proceedings, responsibility may rest with the tribunal or arbitrator controlling the system.11Aceris Law LLC, Who Bears Responsibility for AI in Arbitration? The EU AI Act and the Role of Arbitral Institutions (23 August 2026), discussing the uncertainty surrounding AI use by individual arbitrators, particularly in ad hoc proceedings.
Parties and counsel may also be deployers. However, the draft guidelines indicate that systems used by parties or their representatives to prepare submissions do not fall within the corresponding classification for systems used by or on behalf of judicial authorities.12Draft Guidelines, para 416. This supports an important distinction between using AI to advocate a party’s case and using it to assist in deciding the dispute. Other provisions of the Act may nevertheless apply to parties and counsel.
Finally, once the high-risk regime becomes applicable, Article 25 may also shift provider responsibilities to another actor. It provides, among other things, that a distributor, importer, deployer or other third party will be considered a provider of a high-risk system if it places its name or trademark on the system, makes a substantial modification to it or changes its intended purpose in a way that causes it to become high-risk.13AI Act, Art 25(1). An arbitral institution that merely licenses an existing tool will not necessarily become its provider. The position could change, however, if it materially adapts the system or offers it under its own name for use in arbitral decision-making.
Responsibility will depend on the particular system and how it is used. The Commission’s guidelines remain in draft form, and the institutions’ joint letter is a stakeholder submission rather than an authoritative interpretation. Until the position is clarified, institutions and tribunals should record who selects, controls and operates any AI system used in proceedings.
The AI Act’s Territorial Reach Beyond the EU
Article 2 establishes three territorial connections that may be relevant to an arbitration seated outside the EU, including in Singapore.
First, Article 2(1)(a) applies to providers that place AI systems or general-purpose AI models on the EU market, or put AI systems into service in the EU, “irrespective of whether those providers are established or located within the Union or in a third country”.14AI Act, Art 2(1)(a). An Asian technology company offering an AI tool on the EU market may therefore be subject to the Act even if the system was developed and is operated outside Europe.
Second, Article 2(1)(b) applies to deployers of AI systems that “have their place of establishment or are located within the Union”.15AI Act, Art 2(1)(b). This may capture an EU-based party, law firm, arbitral institution or arbitrator using an AI system in connection with a Singapore arbitration. The relevant obligations would attach to that particular deployer and use, rather than to the arbitration as a whole.
Third, Article 2(1)(c) applies to providers and deployers established or located in a third country “where the output produced by the system is used in the Union”.16AI Act, Art 2(1)(c). For example, this may be relevant where an Asian provider or deployer produces AI-assisted analysis or other material for use by an EU-based participant.
A mixed-location tribunal illustrates the difficulty. Suppose a Singapore-seated tribunal comprises arbitrators working from France, Singapore and Switzerland, while the AI provider is established in the United States. An arbitrator in France who independently uses the system may fall within Article 2(1)(b), while the US provider may fall within Article 2(1)(c) if its output is used in France. This would not necessarily bring the other arbitrators within the Act.
The analysis becomes less certain if the tribunal collectively approves the tool or relies on its output during deliberations. The Act contains no arbitration-specific rule for this situation, so control over the system, the use made of it and where its output is received will matter.
Possible recognition or enforcement of an award in the EU should not, without more, be treated as use of the system’s output in the EU. A Singapore seat therefore does not determine whether the Act applies: Article 2 must be applied separately to each provider, deployer and use of an AI system.
Which Uses of AI in Arbitration May Be High-Risk?
Not every use of AI in an arbitration is high-risk. Classification depends principally on the system’s intended purpose and the function it performs, rather than simply on the fact that it is used during arbitral proceedings.
Under Article 6(2), systems falling within the categories listed in Annex III are to be considered high-risk, subject to the exceptions in Article 6(3). For arbitration, the relevant category is Annex III, point 8(a). It covers systems intended to be used by or on behalf of a judicial authority to assist in “researching and interpreting facts and the law and in applying the law to a concrete set of facts”, and expressly extends to systems intended “to be used in a similar way in alternative dispute resolution”.17AI Act, Art 6(2) and Annex III, point 8(a).
The Commission’s draft guidelines indicate that arbitration may fall within this category where the outcome produces legal effects for the parties and the system assists the decision-making body in interpreting facts or law or applying the law to the case. Arbitral awards will generally satisfy the first requirement because they are binding and capable of recognition and enforcement.18Draft Guidelines, paras 428–431; Convention on the Recognition and Enforcement of Foreign Arbitral Awards 1958, Arts III–V.
The principal distinction is between adjudicative and ancillary uses. Systems that analyse evidence and submissions, draw legal conclusions or generate draft decisions or reasoning are likely to fall within point 8(a). By contrast, tools used for transcription, case management, metadata extraction, document organisation, basic searches or language editing will generally fall outside the category or may qualify for the exception in Article 6(3).19Draft Guidelines, paras 421-423 and p 135, under “Judicial decision/judgment drafting systems”.
Article 6(3) excludes an Annex III system where it does not pose a significant risk to health, safety or fundamental rights and performs one of four specified functions, such as a narrow procedural or preparatory task. Administrative and information-retrieval tools are therefore less likely to be high-risk than systems used to analyse evidence, apply the law or generate draft reasons. An Annex III system that performs profiling of natural persons must, however, always be considered high-risk.
Practical Implications for Arbitrations Seated in Singapore
Apart from regulatory penalties, the most direct risk to an award may arise at the enforcement stage. Article V(2)(b) of the New York Convention permits an enforcing court to refuse recognition where enforcement would be contrary to its public policy.20Convention on the Recognition and Enforcement of Foreign Arbitral Awards 1958, Art V(2)(b). A party resisting enforcement in an EU Member State might argue that non-compliance with the AI Act engages that State’s public policy, even where the award was rendered in Singapore.21Case C-126/97 Eco Swiss China Time Ltd v Benetton International NV [1999] ECR I-3055, paras 36–39. In that case, the Court of Justice held that a fundamental provision of EU competition law could constitute public policy under the New York Convention. That argument remains untested, however, and it does not follow that every breach of the AI Act would meet the generally narrow public-policy threshold. The argument would also require the AI Act to have applied to the relevant use of AI in the first place: the prospect of enforcement in the EU does not itself bring that use within Article 2 of the AI Act.
Even if non-compliance with the AI Act does not itself engage public policy, the AI use in question may still create difficulties under the ordinary rules governing arbitral procedure. Article V(1)(b) of the New York Convention applies where a party was unable to present its case, while Article V(1)(d) applies where the arbitral procedure was not in accordance with the parties’ agreement or, failing agreement, the law of the seat. Undisclosed AI use could become relevant if, for example, the tribunal relied on AI-generated material that the parties had no opportunity to address or used AI contrary to an agreed procedural order. In Singapore, section 31 of the International Arbitration Act gives effect to the corresponding grounds for refusing enforcement of foreign awards.22New York Convention, Arts V(1)(b) and V(1)(d); International Arbitration Act 1994, ss 31(2)(c) and 31(2)(e).
Similar considerations apply where a Singapore-seated award is challenged at the seat. Section 24(b) of the IAA permits an award to be set aside for a breach of natural justice that prejudiced a party. Article 34(2) of the Model Law also applies where a party was unable to present its case, the procedure departed from the parties’ agreement or the award conflicts with Singapore’s public policy.23International Arbitration Act 1994, ss 3 and 24(b), read with the First Schedule; UNCITRAL Model Law on International Commercial Arbitration, Arts 34(2)(a)(ii), 34(2)(a)(iv) and 34(2)(b)(ii). AI use suggesting that the tribunal had delegated its decision-making function rather than exercised its own judgment could also give rise to concern. The relevant ground would be the resulting procedural or public-policy defect, not the breach of the AI Act itself.
An inaccurate AI-generated output or an error in the tribunal’s reasoning would not, without more, satisfy these grounds. The applicant must still identify the rule of natural justice that was breached, show how the breach was connected with the making of the award and establish resulting prejudice.24Soh Beng Tee & Co Pte Ltd v Fairmount Development Pte Ltd [2007] 3 SLR(R) 86 at para 29; ONI Global Pte Ltd v GNC Holdings LLC [2026] SGCA(I) 3 at paras 119–120 and 127. The threshold may nevertheless be met where the tribunal’s use of AI prevented a party from addressing material relied upon in the award or where it is a clear and virtually inescapable inference that the tribunal failed to apply its mind to an essential issue.25CVV and others v CWB [2023] SGCA(I) 9 at paras 32-35; AKN and another v ALC and others and other appeals [2015] SGCA 18 at para 46.
The practical response is to address material AI use early. A tribunal may consider dealing in Procedural Order No. 1 with AI-generated evidence, confidentiality and any disclosure needed to preserve procedural fairness. Targeted directions are likely to be more useful than a blanket requirement to disclose every use of AI. Parties and counsel should likewise identify any relevant EU connection and comply with the tribunal’s directions, even where those directions go beyond the requirements of the AI Act.
Unlike the EU, Singapore has so far approached AI governance principally through existing legislation and voluntary guidance, such as the Infocomm Media Development Authority’s Model AI Governance Framework for Agentic AI. 26Infocomm Media Development Authority, Artificial Intelligence in Singapore; Infocomm Media Development Authority, Model AI Governance Framework for Agentic AI (version 1.5, May 2026). Nevertheless, the EU regime matters to Singapore’s arbitration community. Singapore’s position as a leading international arbitration hub depends partly on the confidence of parties from different regulatory systems and the enforceability of Singapore-seated awards abroad. Even where the AI Act does not directly apply, addressing material AI use through proportionate procedural safeguards is consistent with that confidence.
Endnotes
| ↑1 | AI Act, Arts 4, 5 and 113; European Commission, Timeline for the Implementation of the EU AI Act. |
|---|---|
| ↑2 | AI Act, Art 50(2); European Commission, Timeline for the Implementation of the EU AI Act, above n 1. |
| ↑3 | AI Act, Arts 4, 50 and 99(4)(g) and (6); European Commission, Guidelines on Transparency Obligations for Providers and Deployers of Certain AI Systems, 20 July 2026. |
| ↑4 | Regulation (EU) 2026/1744, Article 1(40)(b)(i), amending Article 113 of Regulation (EU) 2024/1689; OJ L, 2026/1744, 24 July 2026. |
| ↑5 | AI Act, Arts 8–16 and 26, particularly Arts 14 and 26(6); Regulation (EU) 2026/1744, Art 1(40)(b)(i). |
| ↑6 | European Commission, AI Act Service Desk, Frequently Asked Questions – Digital Omnibus: “What are the new timelines proposed?”, stating that the revised timetable acknowledges delays in standards and other implementation support tools. |
| ↑7 | AI Act, Art 3(3) and (4). |
| ↑8 | AI Act, Annex III, point 8(a). |
| ↑9 | AI Act, Annex III, point 8(a); European Commission, Draft Guidelines on the Classification of High-Risk AI Systems, 19 May 2026 (“Draft Guidelines”), paras 428–429. |
| ↑10 | Joint Submission of International Arbitration Institutions in Response to the Targeted Consultation on the Draft Commission Guidelines, 23 July 2026; LCIA, LCIA Leads Cross-Institutional Response to Draft EU AI Act Guidelines, 6 August 2026. |
| ↑11 | Aceris Law LLC, Who Bears Responsibility for AI in Arbitration? The EU AI Act and the Role of Arbitral Institutions (23 August 2026), discussing the uncertainty surrounding AI use by individual arbitrators, particularly in ad hoc proceedings. |
| ↑12 | Draft Guidelines, para 416. |
| ↑13 | AI Act, Art 25(1). |
| ↑14 | AI Act, Art 2(1)(a). |
| ↑15 | AI Act, Art 2(1)(b). |
| ↑16 | AI Act, Art 2(1)(c). |
| ↑17 | AI Act, Art 6(2) and Annex III, point 8(a). |
| ↑18 | Draft Guidelines, paras 428–431; Convention on the Recognition and Enforcement of Foreign Arbitral Awards 1958, Arts III–V. |
| ↑19 | Draft Guidelines, paras 421-423 and p 135, under “Judicial decision/judgment drafting systems”. |
| ↑20 | Convention on the Recognition and Enforcement of Foreign Arbitral Awards 1958, Art V(2)(b). |
| ↑21 | Case C-126/97 Eco Swiss China Time Ltd v Benetton International NV [1999] ECR I-3055, paras 36–39. In that case, the Court of Justice held that a fundamental provision of EU competition law could constitute public policy under the New York Convention. |
| ↑22 | New York Convention, Arts V(1)(b) and V(1)(d); International Arbitration Act 1994, ss 31(2)(c) and 31(2)(e). |
| ↑23 | International Arbitration Act 1994, ss 3 and 24(b), read with the First Schedule; UNCITRAL Model Law on International Commercial Arbitration, Arts 34(2)(a)(ii), 34(2)(a)(iv) and 34(2)(b)(ii). |
| ↑24 | Soh Beng Tee & Co Pte Ltd v Fairmount Development Pte Ltd [2007] 3 SLR(R) 86 at para 29; ONI Global Pte Ltd v GNC Holdings LLC [2026] SGCA(I) 3 at paras 119–120 and 127. |
| ↑25 | CVV and others v CWB [2023] SGCA(I) 9 at paras 32-35; AKN and another v ALC and others and other appeals [2015] SGCA 18 at para 46. |
| ↑26 | Infocomm Media Development Authority, Artificial Intelligence in Singapore; Infocomm Media Development Authority, Model AI Governance Framework for Agentic AI (version 1.5, May 2026). |

