How the EU AI Act applies to arbitration, mediation and dispute resolution

The short version

• The EU AI Act can treat AI that helps decide an arbitration or other ADR whose outcome has legal effects as high-risk, from 2 December 2027 (Annex III point 8(a); recital 61; date set by Regulation (EU) 2026/1744 art 1 point (40)(b)).
• Most mediation sits outside it, because a mediated settlement has no legal effects until the parties agree (draft Commission guidelines, not adopted, paras 430, 431).
• Duties fall on the AI’s provider and its deployer, not on the parties or their lawyers (arts 3, 16, 26).
• Sixteen arbitral institutions asked the Commission not to treat their administrative tools as high-risk (joint letter, 23 July 2026).
• Most support tools, such as search, transcription, scheduling and drafting aids a person finishes, are exempt under Article 6(3) or count as ancillary administration (art 6(3); recital 61).
• A provider or deployer outside the EU is caught where the system’s output is used in the Union (art 2(1)(c)); the AI-literacy and chatbot-transparency duties (arts 4, 50) already apply.^1,3,4,6

Does the EU AI Act apply to arbitration, mediation and other dispute resolution?

It can, but mainly to AI that helps decide a case whose outcome has legal effects. The EU AI Act, Regulation (EU) 2024/1689, can classify such AI as high-risk under Annex III point 8(a), which covers AI used to help a court research and interpret facts and law and apply the law to the facts, “or to be used in a similar way in alternative dispute resolution” (ADR) (art 6(2); Annex III point 8(a)). This guide reads the Act across arbitration, mediation and ADR. The arbitration-only reading is in our guide to AI in arbitration, and the ADR-wide guide is AI in dispute resolution.^1,2

Recital 61 limits the ADR branch to outcomes that have legal effects for the parties, and leaves out purely ancillary administrative activities such as anonymisation, communication between staff and administrative tasks. It also says AI can support a judge but “the final decision-making must remain a human-driven activity”.¹

The draft Commission guidelines of 19 May 2026, a consultation draft that is not adopted, read point 8(a) as two use cases, courts and ADR (para 405). “In similar ways” means the system is used for the same purposes as when a court uses it (para 427), and ADR covers arbitration, mediation, conciliation, ombudsmen and complaints boards (paras 426, 429). Europe-wide court rulings, national court rules and the Council of Europe are in our guide to AI in dispute resolution in Europe.⁴

When is AI used in dispute resolution high-risk?

AI used in dispute resolution is high-risk when the ADR outcome has legal effects and the AI helps research or interpret facts and law, or apply law to the facts. The draft guidelines set a two-part test: first, an ADR body uses the system in proceedings whose outcome produces legal effects for the parties; and second, the system’s purpose is to assist in researching and interpreting facts and law and in applying law to the facts (para 425).⁴

Whether an outcome has legal effects turns on national law, and is decided case by case (para 431). Arbitral awards “generally have binding and final effect”, are enforceable under national law, and are recognised across borders under the New York Convention of 1958, so AI that helps decide an arbitration is the clearest case inside point 8(a).⁴

The draft gives examples that are high-risk (printed pages 135 to 136): AI that analyses the facts and submissions, finds the law and drafts a decision or its reasoning; AI that drafts small-claims decisions or payment orders from structured data and templates; AI that clusters similar cases and proposes text modules for the decision; and AI that identifies the relevant law and precedents and suggests how they apply. The draft lists these for courts, then says the ADR examples follow from them (para 434). Whether such an award can be recognised or enforced is in our guide to AI and award enforcement.⁴

Which AI uses are not high-risk?

Search, transcription, scheduling and other admin tools fall outside point 8(a) altogether, and tools that only do a narrow or preparatory task are exempt under Article 6(3). Under Article 6(3) an Annex III system is not high-risk where it “does not pose a significant risk of harm”, including by not materially influencing the outcome, and it only does one of these: a narrow procedural task (a), improves a previously completed human activity (b), detects decision patterns or deviations without replacing human review (c), or does a preparatory task (d). It never applies to a system that profiles people.¹

• Outside point 8(a), on the draft reading: speech-to-text and transcription; chatbots and other communication with the public; assigning cases; processing and managing evidence; and ancillary administration such as anonymisation (printed pages 136 to 138).
• Exempt under Article 6(3), on the draft reading: pre-classifying incoming claims, extracting metadata, advanced search, language editing, and flagging points a draft has not addressed (printed pages 138 to 139).⁴

Relying on the exception still costs paperwork. A provider that treats its tool as not high-risk must document the assessment before the tool goes on the market, and must register it under Article 49(2) (art 6(4)). How to weigh a tool before buying it is in our guide to AI legal tech.¹

Who is caught: the tribunal, the institution, or no one?

Duties fall on the AI’s provider and its deployer. An arbitrator or tribunal that uses a tool is usually its deployer, and the parties and their lawyers are outside point 8(a). A provider develops an AI system and places it on the market or puts it into service “under its own name or trademark” (art 3(3)). A deployer uses one “under its authority”, except in a personal, non-professional capacity (art 3(4)).¹

The line can move. Anyone who puts their own name or trademark on a high-risk system, makes a substantial modification to it, or repurposes a system so that it becomes high-risk, is treated as its provider (art 25(1)). So an institution or an arbitrator that builds or re-badges its own tool counts as a provider, with a provider’s heavier duties, even while it also deploys the tool.¹

Parties and their lawyers do not act on behalf of the tribunal, so their own AI use is outside point 8(a). The draft says the same about attorneys filing documents before a court (para 416).⁴

The open debate is about institutions. The draft lists “commercial arbitration institutions” as ADR bodies (para 429). On 23 July 2026, 16 institutions asked the Commission to confirm that the high-risk label attaches to the AI a tribunal uses in its adjudicative function, not to an institution’s administrative and organisational tools, pointing to the judicial-administration carve-out (draft paras 411, 412; recital 61). This is the institutions’ position and the draft’s wording, not settled law. The fuller account is in our guide to AI in dispute resolution in Europe, and the joint letter is in our tracker.^4,6

What must providers and deployers do?

A provider of a high-risk system carries the heaviest duties. A deployer has fewer, but must keep a competent person in oversight, retain logs and tell the people a system decides about. A compliance checklist starts with one question: know which role each tool puts you in, then meet that role’s duties.¹

At the demanding end, the Section 2 requirements apply: a quality management system (art 17), technical documentation (arts 11, 18), automatic logging (art 19), a conformity assessment (art 43), an EU declaration of conformity, CE marking and registration (art 49(1)), all under Article 16.¹

A deployer under Article 26 must use the system per its instructions (26(1)), assign human oversight to competent, trained persons who have the authority to act (26(2)), keep the logs it controls for at least six months (26(6)), and inform the natural persons that they are subject to the high-risk system where it makes or assists a decision about them (26(11)).¹

A deployer that is a public-law body, or a private entity providing a public service, must also carry out a fundamental-rights impact assessment (FRIA) under Article 27(1) before it uses a high-risk Annex III system. It may include cross-references to its data-protection impact assessment (art 27(4), as replaced by Regulation (EU) 2026/1744, art 1 point (13)(a)).^1,3

Penalties are set as maxima. Breach of the provider (art 16), deployer (art 26) or transparency (art 50) duties can bring fines up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher (art 99(4)). A prohibited practice carries up to EUR 35 million or 7% (art 99(3)). Member States set and notify the rules (art 99(1), (2)). The high-risk ADR regime does not apply until 2 December 2027, so no fine can yet fall on an ADR body under it.¹

One duty binds everyone already. AI literacy of staff under Article 4 falls on all providers and deployers, high-risk or not. Confidentiality, privilege and data protection when AI touches case material are in our guide to confidentiality and AI in arbitration.¹

Does the EU AI Act reach arbitrations or institutions outside the EU?

It can. Under Article 2(1)(c), a provider or deployer in a third country is covered where the output produced by the AI system is used in the Union. Article 2(1) reaches providers placing AI on the EU market whether or not they are EU-based (a), deployers located in the Union (b), and “providers and deployers of AI systems that have their place of establishment or are located in a third country, where the output produced by the AI system is used in the Union” (c).¹

What “output used in the Union” means for an arbitration is not settled. Candidates include the seat, a party’s location, and enforcement against assets in the EU. The Act fixes no single trigger, so for now a non-EU body cannot read the seat alone as the test. On the widest reading, AI that helps produce an award in a dispute with an EU-seated party, or an award later enforced against assets in the Union, could bring a non-EU body within reach even where the seat is elsewhere (art 2(1)(c)).¹

The practical read for a non-EU institution or neutral: the trigger is use of the system’s output in the EU, not the seat by itself, and two duties do not wait for 2 December 2027. AI literacy (art 4) and chatbot transparency (art 50) already apply. The worked seat, party and enforcement scenarios are in our guide to AI in arbitration; whether an award is enforceable is in AI and award enforcement.¹

When do the rules start, and what already applies?

The high-risk duties for court and ADR AI apply from 2 December 2027. AI literacy and the chatbot-transparency rule already apply. The Act entered into force on 1 August 2024, and its parts start on different dates (art 113, as amended).^1,3

The high-risk date for point 8(a) was once 2 August 2026. The Digital Omnibus on AI, Regulation (EU) 2026/1744, moved it to 2 December 2027. Older write-ups that still give 2 August 2026 for point 8(a) are out of date. The omnibus is adopted and in force since 27 July 2026. It came from proposal COM(2025) 836, and its procedure file, 2025/0359(COD), reads “procedure completed”. It did not delay the prohibitions, which still apply from 2 February 2025, or the Article 50 transparency rule, which still applies from 2 August 2026; it postponed the high-risk Chapter III duties to 2 December 2027 for Annex III and 2 August 2028 for Annex I. The separate data omnibus, COM(2025) 837 (procedure 2025/0360(COD)), is still “awaiting committee decision” and does not touch the court and ADR high-risk regime.^{3,10,11,12,13}

Article 50 is the transparency rule that already bites. Providers must tell people when they interact with an AI system unless it is obvious (50(1)), and must mark synthetic audio, image, video or text as artificially generated (50(2)). Deployers must disclose deepfakes and certain AI-generated text published to inform the public (50(4)). The notice must be clear at the first interaction or exposure (50(5)).¹

The Commission’s own high-risk-guidelines page still describes the classification guidelines as draft, before adoption, and gives the same 2 December 2027 and 2 August 2028 dates.⁵

How does the Act treat mediation and consumer ADR?

AI used in mediation is usually not high-risk, because it falls outside point 8(a): a mediated settlement has no legal effects until the parties agree. Where the outcome is a voluntary agreement or a non-binding recommendation, the draft says an AI system used in it is outside point 8(a) (paras 430, 431). But the Mediation Directive, Directive 2008/52/EC, lets the parties ask a court to make a mediated agreement enforceable, which can change the analysis (para 431; Directive 2008/52/EC art 6). For consumer ADR it is enough that the outcome binds one party, the trader (para 432).^4,8

Consumer ADR has its own rule on top. Directive (EU) 2025/2647 requires consumer ADR bodies that use automated decision-making (automated means in the decision-making process) to inform the parties in advance and let them ask for a review by a natural person. National rules apply from 20 September 2028, and the draft says these safeguards run “in parallel with the AI Act” (para 433).^4,7

For context, the EU closed its own online dispute resolution (ODR) platform: under Regulation (EU) 2024/3228, new complaints stopped on 20 March 2025. Deeper mediation coverage is in our guide to AI in mediation.⁹

What is still open?

• What “output used in the Union” means for an arbitration: the seat, a party’s location, or enforcement.
• Whether the Commission adopts the draft guidelines unchanged, and whether institutions’ administrative tools stay outside point 8(a).
• Who is the deployer in an ad hoc arbitration with no institution, and how the deployer duties (arts 26(11), 27) fall on a sole arbitrator or a small private body.
• Whether an AI-assisted or AI-made award can be refused enforcement, which our guide to AI and award enforcement tracks.
• When a mediated settlement, or an expert determination (a binding finding by a neutral on a defined point), has enough legal effect to fall inside point 8(a).^1,4,5,6