AI in International Arbitration — Benefits, Risks and the Emerging Rules of the Road

Generative AI has moved from novelty to daily tool in international arbitration, and the profession is now past the question of whether to use it and onto the harder question of how. Used well, it compresses cost and time on the mechanical parts of a case and frees lawyers for judgment. Used carelessly, it introduces fabricated authority, leaks privileged material, or taints the process in ways that ultimately threaten the enforceability of the award. This guide maps the benefits against the risks, works through the enforceability and ethical dimensions, and sets out the governance that is beginning to frame responsible use.

1. Scope — what we mean by AI here

The focus of this guide is the current generation of AI, and generative AI in particular — large language models that produce text, summarise, translate and answer questions, alongside the machine-learning tools long used in document review. The relevance to arbitration is twofold: AI as a tool used by parties, counsel, tribunals and institutions to run the process more efficiently; and AI as a subject of the dispute or of new procedural and ethical questions. This guide is chiefly about the first — the responsible use of AI within the arbitral process — because that is where the immediate risks to the award lie.

2. Where AI genuinely helps

• Document review and disclosure. Clustering, de-duplication, relevance triage and privilege screening across large record sets, at a fraction of the manual cost and often with greater consistency.
• Legal and factual research. Rapid first-pass research, issue mapping and synthesis of long records — always to be verified against primary sources before it is relied on.
• Transcription and translation. Real-time hearing transcripts and working translations, particularly valuable in the UAE's bilingual Arabic–English environment.
• Drafting support. First drafts of chronologies, summaries, indices and routine correspondence for a lawyer to refine — never to file unread.
• Arbitrator and counsel research. Analysis of publicly available material to inform appointments, challenges and strategy.
• Analytics and forecasting. Pattern analysis to inform case strategy, cost budgeting and settlement decisions.

The common thread is that AI is strongest on volume and pattern — the tasks that are laborious rather than judgmental — and weakest where the work requires legal judgment, which remains the lawyer's.

3. The risks that reach the award

The dangers are not merely reputational; several bear directly on the validity and enforceability of the award.

• Hallucination. Confident but fabricated citations, quotations and authorities. Filing them is a professional and ethical failure, can mislead the tribunal, and has already drawn judicial censure in litigation.
• Confidentiality and privilege. Feeding privileged or confidential material into a public model may breach confidentiality undertakings and cannot be undone once the data leaves your control.
• Data protection. Processing personal data through AI tools engages data-protection law, including on cross-border transfer — an obligation independent of confidentiality.
• Cybersecurity. New tools widen the attack surface around sensitive dispute data and introduce new vectors for compromise.
• Bias. Models trained on skewed data can embed bias into research, prediction or decision support in ways that are hard to detect.
• Due process. If a party is denied a fair opportunity to address AI-derived material, or if an arbitrator effectively delegates the decision to a tool, the integrity of the process is compromised.

4. The enforceability question

The commercial point that concentrates minds is enforcement, and this is where the abstract risks become concrete. Under the New York Convention, recognition or enforcement of an award may be refused where a party was unable to present its case, where the composition of the tribunal or the arbitral procedure departed from the parties' agreement, or on public-policy grounds. Improper AI use can be mapped onto each of these:

An award that rests on undisclosed AI analysis which the parties had no opportunity to test can be attacked as a denial of the right to be heard. An award in which the arbitrator did not truly make the decision — or used AI in a way the parties had agreed to exclude — can be attacked on composition or procedure. And egregious misuse could be framed as offending public policy. Responsible use is therefore not a compliance nicety; it protects the value of winning, which is the enforceable award.

5. The legal framework — lex arbitri and place of enforcement

Two bodies of law frame the use of AI in any given arbitration. The lex arbitri — the law of the seat — governs the conduct of the proceedings and the standards (fair hearing, equal treatment, tribunal impartiality) against which AI use will be judged, and supplies the grounds on which an award may be set aside at the seat. The law of each likely place of enforcement then governs whether the resulting award will be recognised there. Because an award may need to be enforced in several jurisdictions, the prudent course is to hold AI use to the most demanding standard likely to be applied, not the most permissive.

6. The arbitrator's non-delegable mandate

An arbitrator is appointed to exercise personal judgment, and that mandate cannot be delegated to a machine. AI may organise the record, surface issues or stress-test reasoning, but the decision, its reasons and the assessment of the evidence must remain the arbitrator's own. There is a related transparency dimension: where an arbitrator uses AI in a manner the parties would not reasonably expect, or in a way that could bear on the integrity of the deliberations, disclosure is the safe course. The line to hold is assistance, not substitution.

7. Ethical duties of counsel

For counsel, existing professional duties already govern AI use, even before any AI-specific rule. The duties of competence, candour to the tribunal, confidentiality and diligence all apply directly: competence requires understanding a tool's limits; candour forbids filing unverified AI output; confidentiality forbids exposing privileged material to a public model; and diligence requires human review of everything relied on. A practitioner who treats AI output as a finished product rather than a first draft has not discharged these duties, whatever the tool produced.

8. The emerging rules of the road

Soft-law guidance is crystallising quickly. The Chartered Institute of Arbitrators (CIArb) and the Silicon Valley Arbitration and Mediation Center (SVAMC) have each published guidance on the use of AI in arbitration, and arbitral institutions are issuing notes of their own. Although the instruments differ in detail, their common threads are consistent: agree the ground rules early, ideally at the first case-management conference; verify AI output against primary sources; protect confidential and personal data; keep a human decision-maker accountable; and disclose AI use where non-disclosure could bear on the integrity of the process. Expect procedural orders to address AI expressly as a matter of course.

Bias, explainability and the reasoned award

A specific tension deserves attention: the duty to give a reasoned award sits uneasily with the opacity of many AI systems. An award must be supported by reasons the parties and any supervising court can follow. If an arbitrator were to rely on an AI system whose workings cannot be explained — a 'black box' recommendation as to who should win or how much is due — the reasons would either be the arbitrator's own (in which case the AI has not decided anything) or the machine's (in which case they cannot properly be explained or defended). This is why the safe use of AI in the decision-making phase is confined to transparent, checkable assistance, and why explainability, not merely accuracy, is the quality that matters when AI touches the award.

Institutional practice and procedural orders

Arbitral institutions are responding, and the working assumption should be that AI use will increasingly be addressed in the procedural framework of a case rather than left to individual discretion. In practice this means the case-management conference is the moment to raise AI: to agree whether and how parties may use it, whether disclosure is required, how confidentiality and data protection are to be handled, and what happens if AI-generated material is found to be unreliable. Recording that agreement in a procedural order gives the tribunal a basis to police it and gives the parties certainty. A party that wants to constrain the other side's use of AI — or to protect its own position — is far better placed raising it at the outset than objecting after the event.

A worked illustration

Consider a party that uses a generative model to draft its submissions and, unnoticed, the model invents two supporting authorities that do not exist. If the error is caught by the tribunal or the opponent, the immediate consequences are reputational and may sound in costs; the deeper risk is that the tribunal's confidence in that party's entire case is undermined. Now vary the facts: the party feeds the opponent's confidential disclosure into a public model to summarise it. That is a breach of the confidentiality regime and, if the documents contain personal data, of data-protection law — and it cannot be undone. Both failures are avoidable by the same simple discipline: verify every output against the source, and never expose confidential or personal material to a tool that does not contractually protect it.

Document review and technology-assisted review, in depth

The most mature and least controversial use of AI in arbitration is in the review of documents. Technology-assisted review — predictive coding trained on a sample set, clustering of similar documents, near-duplicate detection and email threading — has been used in large disclosure exercises for years, and the newer generative tools add rapid summarisation and question-answering across a corpus. Used properly, these tools do not replace legal judgment about relevance and privilege; they prioritise and organise the material so that judgment can be applied efficiently.

The safeguards are well understood. The training and the results should be validated and quality-controlled; privilege screening must be robust, because an inadvertent disclosure of privileged material is hard to retrieve; and the process should be transparent enough that, if challenged, the party can explain and defend how the review was conducted. In cross-border matters the location of the review platform and the data raises the data-protection questions addressed below.

Data protection and cross-border transfer

Almost every arbitration involves personal data — names, contact details, financial and employment information — and feeding that data through AI tools is a processing activity that engages data-protection law in its own right. This is an obligation distinct from confidentiality: a party can keep information confidential as between the parties and still breach data-protection law in how it processes personal data. The issues that recur are the lawful basis for processing, the security of the tool and its provider, and, critically, the transfer of personal data across borders to wherever the tool's servers and reviewers sit. Enterprise arrangements with contractual data-protection terms, data-processing agreements and appropriate transfer mechanisms are the norm for responsible use; public consumer tools, which may retain and reuse inputs, are generally unsuitable for anything sensitive.

Cybersecurity and information governance

AI tools widen the attack surface around a dispute at exactly the point when the information is most sensitive. Sound practice mirrors the information-security protocols already familiar in international arbitration: assess the sensitivity of the data at the outset; agree proportionate security measures with the other side and the tribunal; restrict access on a need-to-know basis; use vetted, contractually-bound providers; and plan for incident response. The tribunal can be asked to make directions on information security, and increasingly does. The point is that adopting an AI tool is an information-governance decision, not merely a productivity one, and it should be made with the security of the record in view.

Who is responsible — party, counsel, tribunal, institution

Responsible use is a shared enterprise, and it helps to be clear about who owns what.

• The party decides, with its lawyers, whether and how to use AI, and bears the consequences of misuse in its case.
• Counsel carry the professional duties — competence, candour, confidentiality, diligence — and must verify output and protect data.
• The tribunal controls the procedure, may set the ground rules for AI use, must keep its decision its own, and should be transparent about any use that could bear on integrity.
• The institution sets the administrative framework, may issue guidance, and safeguards the data it holds.

The regulatory trajectory

The governance picture is still forming, but its direction is discernible. Soft-law guidance from professional bodies is consolidating around a common core — disclosure, verification, data protection, human accountability — and institutions are beginning to reference AI in their notes to parties and model procedural orders. Alongside this, the wider legal regulation of AI is developing rapidly in several jurisdictions, and some of it will reach the way disputes are conducted. The practical implication for parties is to build habits now that will satisfy the more demanding standard likely to prevail: assume disclosure may be expected, assume the most stringent data rules of any relevant jurisdiction apply, and keep a human demonstrably in charge.

9. A practical protocol

• Raise AI use at the first procedural conference and record the agreed approach in a procedural order.
• Use enterprise-grade tools with contractual data protection; never paste privileged or personal material into public models.
• Verify every citation, quotation and authority against the primary source before it is relied on or filed.
• Keep a human author and a human signatory accountable for every submission.
• Protect confidentiality and comply with data-protection obligations, including on cross-border transfer.
• Disclose AI use where non-disclosure could later be characterised as a procedural irregularity or a denial of the opportunity to be heard.

Frequently asked questions

Related guides

• The arbitrateAD 2024 Rules — a complete guide
• Arbitration vs litigation in the UAE — when to choose which
• Choosing an arbitration seat in the UAE — onshore vs DIFC vs ADGM

This guide is general information on the law as we understand it and is not legal advice. For advice on a specific arbitration matter, please contact us. Last updated: 30 July 2026.