Inside arbitration series
Access all previous editions for a complete picture of arbitration from our market-leading disputes team
View the full series
21 Oct 2026
Insight
Global
The use of AI in international arbitration is no longer limited to any single task or participant. It now spans document review, legal research, evidence analysis, drafting and, increasingly, the work of tribunals themselves.
The 2025 Queen Mary University of London International Arbitration Survey found that 64% of practitioners already use AI for factual or legal research, and 91% expect to do so within five years.1
That widespread use brings challenges that go beyond any one use case. Much commentary has focused on counsel and experts, but the harder questions are around AI in the hands of the tribunal itself.
In this article, we examine how AI is being deployed across arbitration practice, explore the distinct risks that arise as its use becomes more pervasive, and consider what parties and counsel should be doing now to manage both their own use and the tribunal’s.
How counsel and experts are deploying AI, and how is its usage likely to evolve?
Some applications of AI in arbitration are already largely uncontroversial. Document review and disclosure were an early target: technology-assisted review has over a decade of pedigree, from Moore v Publicis Groupe (SDNY 2012) to Brown v BCA Trading [2016] EWHC 1464, and is expressly contemplated by Article 3.3(a)(ii) of the International Bar Association Rules on the Taking of Evidence. The next generation of AI-powered review tools are similarly becoming commonplace in arbitration.
Unsurprisingly, the equally time-consuming and costly tasks of legal research and drafting are now seeking to draw on the efficiency and quality gains available from AI tools. However, the capability of AI to produce thorough and accurate legal research – particularly in arbitration, where authoritative sources remain paywalled and siloed – is challenging. The direction of travel is nonetheless clear – AI will increasingly ground its results in reliable data and increasingly support lawyers' work across a wide range of arbitration workflows.
The risks of AI in arbitration
As AI becomes embedded across the arbitration workflow, several interconnected risks emerge.
Hallucinations
The most common risk is hallucinations. Courts have admonished lawyers for citing fabricated authorities: in Cork & Anor v Smith [2026] EWHC 1199 (Ch), a firm was criticised for submitting AI-hallucinated legal provisions. Arbitration now has its own example in ARIHQ v Santé Québec (2026 QCCS 1360), where the Québec Superior Court annulled an arbitral award after every citation in the arbitrator’s reasoning turned out to be hallucinated. The court annulled not for the error of law itself – that alone would not have been enough – but because of a serious procedural failing: the entirety of the award’s reasoning was drafted by AI, not the arbitrator. Though a first-instance decision not binding outside Québec, the underlying test – whether the decision-maker ‘drove’ the process rather than outsourcing it to a third party – is presumably the same test tribunals will apply everywhere, just as in the analogous situation of an assistant or tribunal secretary. The principles are unchanged: an AI tool may support an arbitrator, but should never overstep that bright line.
Confidentiality and data leakage
A more practical, everyday risk is potential loss of confidentiality. Feeding submissions, witness evidence or privileged material into public AI tools (i.e., a term used to refer to an AI tool which does not adequately maintain the confidentiality and restricted use of inputs / outputs) exposes the risk of that data being retained, used for training, or being otherwise exposed in a manner that undermines its confidentiality or the privileged nature of the relevant information.
The black box
Arbitration requires an interrogation of the facts, analysis of the law, but most importantly, the exercise of judgment to produce a reasoned award. AI-generated output, by its nature, is not (yet) the exercise of true judgment: it represents the statistically most probable answer, reflecting the biases and limitations embedded in its training data rather than reasoning afresh from the facts and law in their present context. Agreeing with an AI-generated output because it reads plausibly is not the same as exercising independent judgment over the underlying question – and the distinction matters, because a reasoned award requires the latter.
Workflow-level drift
Linked to the above, a further risk arises not from a single wrong answer, but from the quiet accumulation of small skews as AI progresses from discrete tasks to integrated workflows spanning an entire matter. This is compounded by a lack of consistency: unlike a deterministic process, the same AI tool given the same inputs will not necessarily produce the same output twice. In an agentic workflow – where AI is not just answering a single question but executing a chain of tasks, each building on the last – that variability compounds at every step. For example, an AI reviewing documents may systematically underweight a particular category of risk, which then feeds into a skewed case strategy, which in turn shapes the drafting of submissions – no single step may look like an error, but compounded they may cause the case to proceed in a very different and potentially detrimental direction. This risk of compounding AI errors and bias was cited as the greatest deterrent to adoption of AI in international arbitration.2
In truth, many of these risks are not inherent issues with the technology. Rather, they are manifestations of behaviours associated with the use of AI in practice. The risks are manageable, but they point firmly towards well-designed workflows with reinforced human checkpoints throughout – not just a final read-through – run by professionals who understand the system in which they operate. That takes the conversation beyond mere verification: it puts humans in control of the process and maintains their responsibility and accountability for outputs.
The risk clients should be most alive to: when the tribunal reaches for AI
A tribunal’s use of AI is a different beast to the use of AI by counsel, particularly because of the consequences that misuse of AI by a tribunal can have for the enforceability of an award.
In theory, the principle is not new: the question is not who drafted the words, but who is responsible for them and the judgment that underlies them, in the same way the use of tribunal secretaries requires the tribunal to nevertheless own the process and take ultimate responsibility. In the Yukos annulment litigation, the Court of Appeal of The Hague held that it was permissible for a tribunal secretary to have a substantial role in drafting the award provided the arbitrators reviewed it properly and made it genuinely their own; only indiscriminate, unreviewed adoption would cross the line. ARIHQ shows the same test applied to AI: the Québec court drew directly on the arbitrator’s duty to reason independently, likening unchecked AI drafting to letting a law clerk’s work stand in for the decision-maker’s own view. The Chartered Institute of Arbitrators' Guideline on the Use of AI in Arbitration (2025, updated September 2025) makes it explicit: a tribunal must not delegate legal analysis, fact-finding or the application of law to facts to an AI tool where this could influence the outcome. Further useful guidance is also expected from the International Chamber of Commerce Task Force on “Artificial Intelligence in Dispute Resolution” in the form of a protocol on arbitrators’ use of AI, which will build on feedback from the Task Force’s global survey conducted earlier this year.
The risk that parties to arbitration should worry about most is not outright delegation. It is a tribunal using AI to summarise a large volume of submissions or evidence – arguably the most common AI use case – and, in doing so, losing a nuance, a footnote, or the best point in a losing party’s argument, because the summary was never checked against the underlying record closely enough to catch it. This calls into question where exactly a materiality threshold sits, i.e., the point at which AI’s role moves from assistant to arbitrator.
What can parties and counsel do?
1
Be upfront about arbitrator use of AI, ideally expressly addressing it in the procedural order
Ask the tribunal to confirm whether it intends to use AI and for what, and press for a procedural order that confirms the line between routine and high-risk use – disclosure of high-risk use should occur in advance, with an opportunity to comment or object, before the tribunal proceeds. Be equally mindful of where excessive use of AI for routine tasks like summarisation can morph into high-risk use.
2
Ask direct questions
What tool, for what task, how, and what independent verification or workflow safeguards have been implemented?
3
Sense-check more frequently
If reasoning reads as unusually generic, or cites authority that does not check out, challenge it contemporaneously – ARIHQ shows a court will act on exactly that kind of evidence at the set aside/annulment stage if it comes to it. If a tribunal denies using AI and the award suggests otherwise, that discrepancy is itself part of the evidential picture on a challenge and a fact to be established.
None of this requires distrust of arbitrators. It requires the same professional scepticism counsel already bring to any other input into the process.
Practical takeaways from Herbert Smith Freehills Kramer (HSF Kramer)
The answer is not to avoid AI. It is to build the same discipline into our own practice that we ask of tribunals and experts: master the tools, design workflows deliberately, verify output at each stage, and maintain clear records of what AI was used for and how – beyond just a final sense-check. That discipline matters more, not less, as AI shifts from standalone tools to integrated workflows.
AI is part of how our International Arbitration team already operates. But off-the-shelf tools will not always appreciate the factors that matter most in a given dispute, which is why we are redesigning our workflows from the ground up: using AI where it adds genuine value, but combining it with deterministic, code-driven processes where consistency and repeatability matter, and building in features that allow us to audit, check, stage-gate and approve AI-assisted output at each step – not just at the end. Our advice is to be mindful, rather than fearful, of AI, and to ensure that the evidential discipline a dispute demands is embedded in the workflow itself, not bolted on as an afterthought.
Three questions are worth asking on every matter.
- Have we pressure-tested our own AI-assisted workflow(s) on this matter specifically, taking into account the nuances of each matter, rather than assuming a workflow that worked on the last case (or which was created generically) will carry over?
- Do we know if and how our counterparty intends to use AI, and do we have adequate cover in the procedural order that will protect against e.g., the loss of confidentiality, or creation of procedural issues such as hallucinations in their submissions?
- Do we know how the tribunal intends to use AI, and does the procedural order draw a clear line between routine use and anything that could influence the arbitrator’s key decision-making role and ultimately the outcome of the arbitration?
Adoption is not the open question anymore. Discipline is – and that is squarely something that counsel, not just tribunals, must control.
Pages
- Inside Arbitration – Issue 22
- Artificial Intelligence in the hearing room: How AI use is reshaping arbitration practice
- When the pipeline breaks: Force majeure, frustration, and non-performance in energy and commodities contracts
- From the courts to arbitration: The changing shape of FRAND disputes
- Watch this space: Technology, mining innovation and emerging legal risks
- Stacking the deck: Do the new ICC rules make arbitration the winning hand for banking disputes?
- Building the bench: A conversation with Lise Bosman
- Spotlight on Justin D'Agostino
- Building for battle: Navigating construction disputes in the defence sector
Related insights
Legal Notice
The contents of this publication are for reference purposes only and may not be current as at the date of accessing this publication. They do not constitute legal advice and should not be relied upon as such. Specific legal advice about your specific circumstances should always be sought separately before taking any action based on this publication.
© Herbert Smith Freehills Kramer 2026
Stay in the know
Receive timely insights and briefings from HSF Kramer, tailored to keep you informed and ahead
Subscribe now