Lex Arbitri Suite · Writing

AI in arbitration: disclosure is discretionary, duties are not

Disclosure of AI use in arbitration is discretionary, not a default duty. What actually binds sits elsewhere — whether or not the tool is mentioned.

Must the use of artificial intelligence in arbitration be disclosed? An opponent can cast counsel's silence as concealment, or use an arbitrator's admission to challenge the appointment or the award. None of the instruments that now guide the field requires disclosure as a general rule. The duties they do impose sit elsewhere, and they bind whether or not the tool is ever mentioned.

No default duty to disclose

The Silicon Valley Arbitration and Mediation Center published its Guidelines on the Use of Artificial Intelligence in Arbitration in April 2024; the Chartered Institute of Arbitrators followed with its own Guideline in 2025. SVAMC Guideline 3 states in terms that 'disclosure that AI tools were used in connection with an arbitration is not necessary as a general matter', to be decided instead 'on a case-by-case basis'. Its commentary goes further, explaining that a general disclosure rule 'does not seem feasible and could create more controversy than it resolves' (SVAMC Guidelines 2024, Guideline 3).

The CIArb Guideline is readier to contemplate an obligation, but a conditional one: disclosure 'may be required to the extent that its use may have an impact on the evidence, the outcome of the arbitration or otherwise involve a delegation of an express duty toward the arbitrators or any other party', and it is for the tribunal, after consulting the parties, to impose any such obligation (CIArb Guideline 2025, art 7.1, 7.3).

Neither instrument creates a standing, self-executing duty to announce that an AI tool was used. Disclosure is discretionary: it turns on the impact of the use, and where an obligation arises, it is for the tribunal to impose.

The duties that bind

The absence of a disclosure rule is not an absence of duties. AI use carries several; they attach elsewhere.

One objection must be met first: both instruments are soft law. The SVAMC Guidelines apply only where the parties or the tribunal adopt them (SVAMC Guidelines 2024, Preliminary Provisions), and the CIArb Guideline likewise binds no one by its own force. Why treat any of it as binding? Because the duties that matter do not draw their force from the guidelines. Confidentiality is owed under the parties' agreement, the applicable rules, and a practitioner's own professional obligations, whether or not a guideline is adopted; the tribunal's power to regulate the use of AI and to require disclosure exists within its general authority to conduct the proceedings (CIArb Guideline 2025, art 4.1); and an arbitrator's mandate is personal as a matter of the law of the seat, not of a soft-law annex — arbitration statutes appoint persons, require their signatures on a reasoned award, and allow the award to be set aside where the agreed procedure was not followed (UNCITRAL Model Law, arts 11, 31, 34(2), given force by national enactments such as Singapore's International Arbitration Act and Hong Kong's Arbitration Ordinance, and matched in the leading non-Model-Law seat of London by the Arbitration Act 1996, ss 33, 52 and 68). The guidelines do not create these duties. They locate them for a new tool.

The first is confidentiality, and it binds every participant — counsel, arbitrator, expert, and institution alike. SVAMC Guideline 2 makes all participants 'responsible for ensuring their use of AI tools is consistent with their obligations to safeguard confidential information', and tells them in plain terms not to 'submit confidential information to any AI tool without appropriate vetting and authorisation'. Only tools that 'adequately safeguard confidentiality' should be used with protected material, and materials should be redacted or anonymised where appropriate (SVAMC Guidelines 2024, Guideline 2). It is also the duty most easily breached, and the breach completes at once: a privileged witness statement pasted into a consumer chatbot whose terms reserve the right to retain and train on inputs is compromised the moment it is sent. An arbitrator who drafts with the tribunal's deliberations open in the same chatbot is in the same breach. The CIArb Guideline flags the hazard as a first-order risk of the technology (CIArb Guideline 2025, art 2.2).

The second — on parties and their representatives — is accuracy, and it is older than the technology. Counsel have always owned what they file. SVAMC Guideline 4 does not create the duty; it restates it for a tool whose polish invites the drafter to skip the checking: parties 'shall review the output of any AI tool used to prepare submissions to verify it is accurate', and parties and party representatives 'shall be deemed responsible for any uncorrected errors or inaccuracies' (SVAMC Guidelines 2024, Guideline 4). 'Deemed' is the operative word — the same attribution that already makes a solicitor answer for a junior's draft, now denying the machine any special excuse. The instruments do not ask counsel to reveal that they used AI. They ask counsel to stand behind what it produced as their own work — which, in the eyes of the rule, it is.

The arbitrator's harder line

The duties so far govern how the tool is used: handle the parties' information with caution, check its work. For arbitrators the instruments go further — from duties of conduct to a line that must not be crossed. SVAMC Guideline 6 provides that 'an arbitrator shall not delegate any part of their personal mandate to any AI tool', a principle that applies 'particularly' to the decision-making process (SVAMC Guidelines 2024, Guideline 6). The CIArb Guideline is consistent: arbitrators 'should not relinquish their decision-making powers to AI', should avoid delegating legal analysis or the application of law to the facts where that could influence the outcome, and 'shall assume responsibility for all aspects of an award, regardless of any use of AI' (CIArb Guideline 2025, art 8.2, 8.4). The model agreement annexed to the CIArb Guideline puts it without qualification: the arbitrator 'shall not delegate any decision-making responsibilities to the AI Tool' (CIArb Guideline 2025, Appendix A, art 10.2.4).

The absolute language marks a difference in kind. An arbitrator's appointment is intuitu personae — made in the person of the arbitrator. One who uses a model to reason toward a conclusion, rather than to organise material already reasoned through, no longer supplies the judgement the parties contracted for. The instruments name the stake. CIArb ties AI use and its disclosure to 'the integrity of the arbitration and/or the validity and enforceability of the award' (CIArb Guideline 2025, art 7.2), and when UNCITRAL convened a colloquium on AI in dispute resolution at UN Headquarters in February 2026, it framed its arbitrator panel around that stake — 'Arbitrators and the Use of AI – Impact on Recognition and Enforcement' — in a practice where, the programme observed, 'some arbitrators seek party agreement before using AI, others disclose its use in the award, while some do neither' (UNCITRAL WGII Colloquium Programme, February 2026, Panel 1). Delegation puts the award itself at risk.

The risk is no longer prospective. In April 2026 the Superior Court of Québec annulled an award upon finding that the authorities supporting it — a doctrinal article, three court decisions, an arbitral award — did not exist: they were AI hallucinations, and they were the only legal support the award had. The evidence, the court held, led to the conclusion that 'the Arbitrator's authority was delegated and he abdicated his role of reviewing the result'; the award fell for breach of the agreed arbitral procedure, and the parties were ordered to appoint a new arbitrator within sixty days (ARIHQ c Santé Québec, 2026 QCCS 1360, paras 103–124; our translation). The line the court drew is SVAMC Guideline 6's own, though reached independently — from Québec law and the judicial-AI guidance the court itself cited, not from the arbitral instruments: researchers, clerks and citation aids remain unobjectionable, and not every award that cites a wrong reference, or that was drafted with the help of AI, will fall — this one fell because the hallucinated apparatus stood 'at the heart of the Arbitrator's reasoning' (paras 87, 113, 117). A year earlier, a federal petition in California had attacked an award on the same ground — the sole arbitrator, it alleged, had exceeded his powers by 'outsourcing his adjudicative role to artificial intelligence' — and was dismissed for want of jurisdiction without the question ever being reached (LaPaglia v Valve Corp, No 3:25-cv-00833 (SD Cal, 9 December 2025)). One challenge landed and one did not — and together they show what an award now faces. Because no arbitrator admits to letting the machine decide, the allegation has to be proved circumstantially: in Québec from the court's check of every cited authority against the databases, which found none existed; in California from the award's invented facts — findings that appeared nowhere in the record — and the arbitrator's admitted use of ChatGPT. That such a challenge can be run at all is the new thing — and even when it fails, the party who won the arbitration loses a year to defending the award.

A deliberate restraint

The pattern is deliberate. The drafters imposed duties where breaches do harm — confidentiality, accuracy, the mandate — and withheld a disclosure rule that would catch every spell-checker and translation tool while preventing none of those breaches.

Nor was the choice two committees' idiosyncrasy. The Stockholm Chamber of Commerce, in a guide adopted in October 2024 expressly 'without imposing specific obligations', goes no further than encouraging tribunals 'to disclose any use of AI' in researching or applying the law — and states flatly that a tribunal 'cannot delegate the decision or the reasoning leading to it to anyone or anything' (SCC Guide 2024). VIAC leaves it 'within their discretion' whether arbitrators inform the parties of their own use, and whether to request disclosure of evidence produced with AI, while writing non-delegation as 'shall not' (VIAC Note on AI 2025, paras 2.1, 5.2, 6.1). CIETAC makes disclosure a matter for the parties to agree or the tribunal to invite, and reserves its firm words for the tribunal itself, which must not cede its decision-making power to the tool (CIETAC AI Guidelines 2025, paras 3.1, 4.2.2, 5.2; our translation — no official English text is published). A Madrid working group's principles for arbitrators' use of AI contain no disclosure principle at all, and state the mandate in the register the appointment carries: the arbitrator answers intuitu personae and 'is the sole decision-maker' (CIAM-CIAR, Arbitraje e Inteligencia Artificial 2025, principles 5 and 8; our translation). Drafting bodies in Stockholm, Vienna, Beijing and Madrid, none obliged to follow the first two, kept arriving at the same asymmetry.

The clearest evidence that the restraint was chosen is that some courts, facing the same risk, were making the opposite choice at the same time. In May 2023, with sanctions proceedings pending against two New York lawyers over a brief citing cases that ChatGPT had invented, a federal judge in Texas issued a standing order requiring every lawyer before him to certify either that no part of a filing was drafted by generative AI or that a human had checked any AI-drafted passage (Judge Brantley Starr (ND Tex), Mandatory Certification Regarding Generative Artificial Intelligence, 30 May 2023). The following month, the New York court sanctioned the lawyers under Rule 11 and required them to notify every judge they had falsely cited (Mata v Avianca, Inc, 678 F Supp 3d 443 (SDNY 2023)). Judges in other courts followed with orders of their own, and the instinct is not confined to the United States: before the DIFC Courts in Dubai a party may not use AI-generated content in proceedings without first securing the other side's agreement or, failing that, the court's approval, and 'should declare at the earliest possible opportunity' any use of it (DIFC Courts, Practical Guidance Note No 2 of 2023). The colloquium's own courts presentation put that regime before the arbitrator panel, alongside its blunt answer to whether lawyers should disclose AI use to courts: 'current regulations suggest yes' (Sapna Jhangiani KC, 'AI and Courts', UNCITRAL WGII Colloquium, 16 February 2026).

Courts and arbitration are not solving an identical problem. A court answers for public justice and takes every litigant who walks in, including the unrepresented; arbitration is consensual and arrives with a tribunal already empowered to police misuse case by case. But that is what makes the guidelines' restraint readable as a decision: by the time SVAMC published in April 2024, the certification orders were in force and widely reported. Their own commentary weighs a general disclosure rule and rejects it.

The certification instinct has itself since softened. When the Civil Justice Council examined the use of AI in preparing court documents for England and Wales, its interim report provisionally concluded that, provided a statement of case bears the name of the lawyer taking professional responsibility for it, 'there is no need for any (further) rules relating to statements of case produced with the assistance of AI' — reserving declaration requirements for the places where the use touches the evidence itself, trial witness statements and expert reports (CJC, Use of AI for Preparing Court Documents: Interim Report and Consultation (17 February 2026), paras 4.7, 7.15–7.16, 8.2); on the consultation's close in April 2026, the Council's June 2026 summary of responses recorded a high degree of alignment that no AI-specific rule was needed for statements of case, with the final report to follow later in 2026 (CJC, Use of AI in preparing court documents — update). That is the arbitral instruments' architecture reached from the courts' side: responsibility for the work as the rule, disclosure where the evidence is touched.

Where the principle is heading

The UNCITRAL colloquium shows the direction of travel. UNCITRAL convened it during the eighty-third session of Working Group II so that the Working Group could recommend what the Commission should take up next on AI in dispute resolution. Its panels tabled a menu of what UNCITRAL might one day take up — though the Working Group itself judged it premature to decide the form of any future work, recommending only that the secretariat keep monitoring and report back (Report of Working Group II on its eighty-third session, A/CN.9/1237 (4 March 2026), paras 65, 70). The restraint the instruments show, the institution shows too. Among the proposals the report records (para 48), three came from SVAMC's contribution, presented by Professor Crenguţa Leaua. The first extends arbitrator independence to the arbitrator's relationship with the machine, under a name of its own — 'cognitive independence and impartiality': independence defined in the presentation as 'no delegation of reasoning to AI', impartiality as 'protection from algorithmic thinking distortions' (Crenguţa Leaua, 'Proposals for Future Work', UNCITRAL WGII Colloquium, 16 February 2026; recorded in the Working Group's report as addressing 'the subtler risk of algorithmic influence on how arbitrators reason, frame issues, and prioritize information, thereby filling the normative gap in existing doctrines of independence and impartiality' (A/CN.9/1237, para 48)). The point is not only to refuse to outsource the decision but to guard one's own reasoning against a tool that is confident, fluent, and often wrong. The second, 'dynamic due process', asks how a party's consent to arbitrate stays meaningful as the technology shifts under it after the agreement is signed. The third would reserve decisions in equity — ex aequo et bono — as what the presentation calls 'exclusively human' territory, off-limits to the logic of a statistical model. All three are about keeping human judgement at the centre of adjudication as the tools grow more capable of imitating it. The same list put disclosure on the table too — model clauses and procedural-order language on 'disclosure and transparency … the type of tool used, the scope and timing of use, and consequences of non-disclosure' (A/CN.9/1237, para 48): the very rule the current instruments withhold, now among the things UNCITRAL may one day be asked to build.

The American Arbitration Association's International Centre for Dispute Resolution (AAA-ICDR) told the same panel that its AI Arbitrator is 'launched and available for use' — two-party, documents-only construction disputes for now, with explicit party opt-in — a system in which 'AI performs structured legal reasoning, drafts decisions', and a human arbitrator 'reviews, edits if necessary, and issues the final award' (Diana Didia, AAA-ICDR, UNCITRAL WGII Colloquium, 16 February 2026). SVAMC Guideline 6 forbids an arbitrator delegating any part of the personal mandate; the AAA-ICDR system tests that rule directly. When the machine drafts the decision and the human signs it, has the arbitrator decided, or only ratified? The institutions are already working that boundary — the IBA on a 'traffic light' system to sort which tasks across the arbitration life cycle may be delegated to AI, an ICC task force on how arbitrators may use it, alongside the CIArb guidance (A/CN.9/1237, para 46); and the IBA's launch of a dedicated Artificial Intelligence Institute in June 2026, its founding director, Dr Farzana Dudhwala, previously Meta's Global AI Policy & Governance Lead (IBA, 8 June 2026), marks where institutional attention is moving — from governing the tools to governing the act of judging itself: with AI in the room, the human must still be the one who decides.

What to do now

For the practitioner, three rules carry most of the weight:

For an arbitrator the last of these is not advice but the core of the mandate — the tool may organise, summarise, and speed the reading, but the judgement stays with the person the parties chose.

Disclosure is the visible part of the problem, and the smaller part. The instruments that now guide AI in arbitration ask less about what a practitioner reveals than about what they protect and what they refuse to hand over — the parties' confidences, and the judgement at the centre of the decision. Attend to those, and the disclosure question largely answers itself: where the use touches the evidence or the outcome, the tribunal will shape what must be said. Disclosure is discretionary. The duties are not.