AI in mediation: how it is used, what the guidance says, and the limits

The short version

• AI in mediation is mostly support work: preparing, summarising, translating and administering cases, while the parties keep control of the outcome.
• The only dedicated international instrument is the IBA Mediation Committee’s guidelines of 19 June 2025, a non-binding living document that lists uses and sets four safeguards.
• Confidentiality is the hardest problem, above all keeping a private session (a caucus) from leaking to the other side.
• No rule or guidance we read lets AI replace a mediator. A mediated settlement is still help from a human third person.
• The evidence is strongest for admin tasks and weakest for the human side, and no study has compared AI with a trained mediator.
• From 2 December 2027 the EU AI Act can treat AI that helps decide an alternative dispute resolution (ADR) case with legal effects as high-risk, but the Commission’s draft guidelines leave out AI where the outcome is a voluntary agreement, as in most mediation.^{1,2,3,4,5,8,9,22,24,28,29,30}

What is AI in mediation?

AI in mediation is the use of AI tools, mostly generative AI, by the mediator, the parties and their representatives, or the mediation institution running the case. Generative AI means systems that create content from what a user gives them, as a large language model (LLM) does. The IBA guidelines set that scope and leave out non-generative tools that work on fixed rules or algorithms (scope section, p. 2).¹

One use keeps a person deciding: AI that assists a human mediator, a human in the loop. The other is the idea of AI mediation: an AI mediator that conducts the mediation itself. For AI across dispute resolution, including online dispute resolution (ODR), ombudsman schemes and courts’ online services, see our guide to AI in dispute resolution.¹

How is AI used in mediation?

AI in mediation is used mostly to prepare, summarise, translate and administer cases around a human decision. The IBA guidelines list general uses (Part one, section 1) and then group the rest by who is doing them (sections 2 to 4).¹

• Mediators (Part one, 2.1 to 2.8): screening a dispute for suitability, tailoring the procedure, building summaries, timelines and maps of who is involved, analysing positions and power dynamics, preparing questions, real-time summaries where participants agree to transcription (2.6), generating settlement options, and watching for imbalance so parties are treated equally.
• Parties and party representatives (3.1 to 3.6): litigation-risk analysis, choosing a mediator, position and interest papers, preparing for and supporting sessions, and first drafts of settlement agreements, with any AI legal advice reviewed by a legal professional (Part two, 1.2).
• Mediation institutions (4.1 to 4.8): case administration, fees, interpreting unclear dispute resolution clauses, building mediator panels, and attestations for enforcing settlements, including under article 4 of the Singapore Convention.^1,8

Institutions have put some of this into practice. The Singapore International Mediation Centre launched an AI assistant on 28 August 2024 that builds chronologies, maps who is involved, summarises disputes and drafts an initial settlement agreement “for counsel to revise and finalise”. The American Arbitration Association (AAA) launched an AI-powered mediator search on 6 October 2025 and an AI Agreement to Mediate Builder on 19 May 2026. A UN Working Group II report records experimental uses discussed at a 2026 colloquium, including automated note-taking, emotion analysis, creative proposal generation and predictive outcome analysis (A/CN.9/1237, para 50).^4,13,14,15

A 2025 mediation audit found that most mediators who use AI do so for their own summaries and reading time, not with the parties in the room.¹²

What do the IBA guidelines say?

The IBA Mediation Committee’s guidelines of 19 June 2025 are the only dedicated international instrument on AI in mediation. They are non-binding, cover only generative AI, and are built as a list of uses (Part one), safeguards (Part two) and a sample disclosure statement (Part three).^1,2

Disclosure is not mandatory under the guidelines, which say it may be required or helpful, and Part three offers a sample AI usage statement a participant can adapt. The committee calls the text a living document, and as of 30 October 2026 we found no revision. The IBA guidelines in our AI rules tracker.^1,2

Can AI replace a mediator?

No rule or guidance we read lets AI replace a mediator, and the texts that define mediation describe help from a human.³

• UNCITRAL’s secretariat describes an agreement to mediate as an agreement “to seek a human mediator” when it asks whether a mediator may rely on AI (A/CN.9/1215, para 47).
• The Singapore Convention defines mediation as help from “a third person or persons” who lack the authority to impose a solution (art. 2(3)), and the UNCITRAL Model Law on International Commercial Mediation uses the same definition (art. 1(3)). Neither mentions AI.
• States answering UNCITRAL stressed human oversight, that AI “should remain subordinate to human judgment” and that “final decisions must be taken by natural persons” (A/CN.9/1271, para 14). UNCITRAL itself, at its 2026 session, judged the topic “not yet sufficiently mature to warrant legislative work” and kept its AI programme exploratory (A/81/17, paras 170, 172).^3,5,6,8,9

A presentation to the UNCITRAL colloquium in February 2026 reported that commercial platforms marketing AI-avatar or AI-assisted mediation are being offered. No court or regulator we found has ruled on one, and the evidence does not support handing the mediator’s role to a machine (next section).⁷

The wider question, whether AI can decide any dispute, reaches arbitrators and judges too: see can AI replace a mediator or decide a dispute and whether AI can be an arbitrator.⁶

What does the evidence show?

AI performs best on administrative and document work, where the stakes are lowest, and worst on the interpersonal, culturally sensitive parts of mediation. No study has compared AI with a trained, professional mediator.⁴

• A UN Working Group II report says AI tools are “currently most reliable for administrative and document-processing tasks”, that reliability “decreased in analytical tasks and was weakest in interpersonal and culturally sensitive contexts”, with “persistent hallucination risks” and “reduced perceived empathy when parties knew responses were machine-generated”, and that “no research to date compared artificial intelligence use directly with trained mediators” (A/CN.9/1237, para 51).
• A 2024 study in Science trained an AI to help a group find common ground in deliberation; participants (N = 5,734) preferred its statements to those the human mediators in the study wrote. This was group deliberation, not two-party dispute mediation.
• A 2024 study in the Proceedings of the National Academy of Sciences found that AI messages could make people feel heard, and that the effect dropped once they knew the message came from AI.
• A 2025 study in Nature Human Behaviour (nine studies, n = 6,282) found human-attributed responses were rated more empathic and supportive, and that people chose human interaction over AI for emotional engagement.
• A 2024 study in PNAS Nexus found LLMs default to the values of English-speaking and Protestant European countries; cultural prompting improved alignment for 71 to 81% of countries for later models but did not remove the bias.
• Purpose-built legal AI research tools still return false or unsupported answers, a hallucination, 17 to 33% of the time.^{4,22,23,24,25,26}

Is AI safe for confidential and caucus material?

Treat anything put into an AI tool as possibly exposed, and take extra care with the caucus, the private session with one party, because a mediator’s confidentiality duties hold whatever tool is used.¹

• The IBA warns that confidential information entered into LLMs “may be vulnerable to data breaches and unintended disclosure”, and says to anonymise and limit what is entered and to review a tool’s terms first (Part two, 2.1 to 2.3).
• Caucus is the sharp edge. The UN report records the concern that AI systems “might inadvertently disclose information shared in private caucus that should remain confidential from the other party” (A/CN.9/1237, para 51), and a November 2025 seminar reported by UNCITRAL named caucus confidentiality a key challenge (A/CN.9/1271, para 17).
• The baseline ethics duty sits in the Model Standards of Conduct for Mediators (adopted by the American Arbitration Association, the American Bar Association and the Association for Conflict Resolution, 2005): a mediator “shall maintain the confidentiality of all information obtained by the mediator in mediation” unless the parties agree or the law requires otherwise (Standard V(A)), and shall not pass on what is said in a private session “without the consent of the disclosing person” (Standard V(B)). The IBA safeguards add to these duties, they do not replace them.
• Where EU law governs, mediation confidentiality is protected by statute: Directive 2008/52/EC says mediators generally cannot be compelled to give evidence about a mediation (art. 7). Any AI use must not defeat that.^1,4,5,10,11

The deeper data-protection and privilege questions, including who is the controller or the processor and what cross-border transfer needs, are in confidentiality, privilege and data protection when AI is used.¹

Should mediators disclose AI use, and get consent?

Disclosing AI use is not required under the IBA guidelines, which say it may be required or helpful, and consent is expected before AI changes the process, above all before an AI note-taker records the session.¹

• IBA: any participant may ask which AI tools another is using and why, a sample AI usage statement sits in Part three, and if a concern cannot be resolved together the tool “should be discontinued” (Part two, 4.1 to 4.2; Part three).
• The Singapore Mediation Centre’s Prompt Engineering Guide for Mediators tells mediators to be transparent and gives a sample statement (rule 6, p. 4), and the AAA’s AAAi Standards for AI in ADR ask neutrals to tell participants how AI informs their process (Standard 4).
• Consent for recording is the practical test. The IBA ties real-time AI summaries to participants agreeing to transcription (2.6); a Resolution Institute consultation draft (16 September 2025) would require the written informed consent of all participants for AI note-taking, with a right to withdraw; and a Civil Mediation Council draft code asks for consent before any digital tool is used (para 2.8).
• Competence matters too. Under the Model Standards a mediator should act only with “the necessary competence” (Standard IV(A)), which now takes in understanding the AI tools in use.^{1,10,16,17,18,19,20}

How should parties use AI?

Parties and their representatives, including self-represented parties, can use AI to prepare and decide better: risk analysis, position papers, weighing offers and first drafts of settlement terms. They should keep their own judgement and not treat AI as a “final authority” (IBA Part one, 3.1 to 3.6; Part two, 1.1 to 1.2). Mediation rests on party self-determination (Model Standards, Standard I).^1,10

Whatever a party puts into a tool carries the same confidentiality and accuracy risks as it does for the mediator, so the caucus and upload cautions above apply to them too.¹

Does the EU AI Act apply to mediation?

Mostly no. The EU AI Act can treat AI that helps an ADR body decide a case with legal effects as high-risk, but most mediation ends in a voluntary agreement, which the European Commission’s draft guidelines place outside that class.^28,29

• Annex III point 8(a) covers AI used by or for a judicial authority to research and interpret facts and law and apply the law to facts, “or to be used in a similar way in alternative dispute resolution”. Recital 61 limits the ADR part to outcomes with legal effects for the parties and excludes purely ancillary administrative activities.
• The Commission’s draft guidelines of 19 May 2026, not adopted, say AI used where the outcome is “a voluntary agreement or non-binding recommendation” is outside point 8(a) (para 430), and that both conditions, an ADR outcome with legal effects and a research-and-apply purpose, must be met (para 425).
• From 2 December 2027 the Act can treat in-scope AI as high-risk under the Digital Omnibus on AI, Regulation (EU) 2026/1744 (art. 1 point (40)(b)).^27,28,29,30

This is a signpost. The full role-and-duty read, provider versus deployer and the institution-versus-tribunal debate, is in how the EU AI Act applies to arbitration, mediation and other dispute resolution, and Europe-wide court rulings are in AI in dispute resolution in Europe.²⁹

What should mediators and institutions do now, and what is still open?

In practice, keep confidential material and the final decision with people, get written consent before any AI records a session, disclose the tools in use, and sort each tool by whether its output has legal effects. Some questions are still open (below).^1,19,29

• Keep confidential material out of public tools and anonymise what goes in (IBA 2.2; Singapore Mediation Centre rules 2 and 3; WIPO Note section 4).
• Get written consent before AI note-taking (Resolution Institute draft), and participants’ consent before any digital tool is used (Civil Mediation Council draft code, para 2.8).
• Use a disclosure statement, the IBA Part three sample or the Singapore Mediation Centre sample (p. 4), and answer participants’ questions about the tools in use.
• Keep the decision and the settlement with the parties and the mediator.
• Sort each tool by whether its output has legal effects (AI Act recital 61; draft guidelines para 430).^{1,18,19,20,21,28,29}

Some questions are open. Whether a settlement reached on flawed AI advice could be refused enforcement under the Singapore Convention (art. 5(1)(e)) is unsettled (A/CN.9/1215, para 46). So is whether an AI system could issue an attestation that a settlement came from mediation, one of the forms of evidence the Singapore Convention accepts (art. 4(1)(b); A/CN.9/1237, para 52). UNCITRAL has flagged future work on confidentiality, disclosure, accuracy standards and mediator liability (A/CN.9/1237, para 53). We found no institution that has published outcome data on AI-assisted mediation (how we search).^3,4,8