Artificial Intelligence in Arbitration
Whose Decision?
Falsehood flies, and the truth comes limping after it, so that when men come to be undeceived, it is too late; the jest is over, and the tale hath had its effect.
Jonathan Swift
Introduction
On its landing page, Anthropic, the makers of Claude, proclaims the lofty aim of securing AI’s benefits and mitigating its risks. Similarly, OpenAI developed ChatGPT as a charity with its “mission is to ensure that artificial general intelligence benefits all of humanity.”
As I write this, Elon Musk’s $150 billion litigation against OpenAI has failed, not because his claim “it is not okay to steal a charity” had no merit, but because a jury found that the claim was out of time. The issue is important as the intrusion of AI into our professional lives is an inevitability. We will be unable, like Knut, to hold back the tide. To paraphrase the Daleks of a bygone age, resistance is futile!
That said, the enthusiastic and untrammelled adoption of AI, and the most recent word prediction engine the Large Language Model (LLM), is not without its risks, not least in the tendency to invent “facts” and cases (styled hallucinations) with the apparent intent to fill gaps and to please the interrogator. There has been considerable judicial criticism, most recently in February 2026 when the Supreme Court was prompted to comment:
Misuse of AI in legal proceedings has serious implications for the administration of justice and public confidence in the justice system. Persons filing submissions in court must ensure all authorities referred to are genuine and correctly cited.[1]
The reality of artificial intelligence (AI), or generative AI, is that it is here and it’s here to stay. The topic is everywhere, and it is less about the motives of the developers of these technologies than about its impact and how to maximise the benefits it brings, while avoiding or minimising the all too apparent risks.
Hallucinations are, perhaps, the relatively simple manifestation of what will be a wider problem with which counsel, arbitrators and institutions need to come to grips – at what point does AI involvement end and where must human input be preserved?
The Roles of Counsel and Arbitrators
The chronology of a dispute will (should) involve counsel providing initial informed and researched advice, research, advice on the appointment of the arbitral tribunal and preparation and presentation of the case. The tribunal then provides a considered determination of the dispute, based on counsels’ submissions and its own knowledge and skill.
The client is not simply embarking on a process, but is first relying their lawyer’s considered opinion; and second, the arbitral tribunal to provide a robust and durable determination of the dispute. Both stages require critical thinking and the application of professional skill.
In his wonderful 2010 book The Rule of Law, the late Master of the Rolls, Tom Bingham, sets out three duties for lawyers, and I would add by extension, arbitrators –
(1) to be available – at its simplest, don’t accept engagement if you don’t have the capacity to apply yourself to it;
(2) to use your skills and training – again, taking the negative view, don’t accept appointments which are beyond your skills; and
(3) to maintain your professional integrity – it’s not about you, but about professionalism in all its shades of meaning.
Those duties have survived through significant changes in technology and legal practice, from devilling in a law library, drafting advice, formal representation of the case and its determination, in each case applying available technologies which have developed over only 40 years or so from typing through word processing, facsimile machines, the internet, email and now artificial intelligence agents taking the drudgery out of the mundane activities. While those technologies have undoubtedly improved efficiencies, they have also resulted in a significant increase in the volume and complexity of submissions and awards.
The pace of this most recent change is the more startling. Anthropic has released an upgrade to its plug-in, Claude for Legal, in the last two weeks and Harvey has been the AI assistant of choice for legal firms for some time. Both offer options tailored to legal practice. Things are moving fast!
In his 2021 Reith lectures, Stuart Russell, Professor of Computer Science at UC Berkley, warned that artificial intelligence technology was developing without any protocols or ethical limits to where it might lead us. He commented “success would be the biggest event in human history … and perhaps the last event in human history.”
What is not required is a modern version of the 19th century Red Flag Act[2], but a 21st century recognition of the benefits that AI can provide while protecting the critical thinking and professionalism that arbitral proceedings have to offer.
Jevons’ Coal Question
In 1865, English economist William Jevons posited in his book The Coal Question that there was a coal paradox; he predicted that with technology (notably Watts’ steam engine) there would be a reduction in consumption of coal for a given output, with the result that demand for coal would increase exponentially, culminating in “peak coal”. In effect, there would be a shift in the production bottleneck. He was wrong about the exponential nature of coal demand as other technologies, notably oil, intervened. But he was correct about the knock-on effect of improved technologies.
The issue Jevons’ paradox identified is that in any given system, technology will produce greater efficiencies which will shift the natural bottleneck, whether it’s the availability of resources, cheaper production or market access. The critical point is that technologies bring efficiency, which removes or reduces the effect of those bottlenecks, increasing the opportunities for innovation and, in our case, critical thought.
The unintended consequence of most technologies in law has been the explosion in the length and complexity of documentation. That has produced a bottleneck of its own – human resources to manage, assess and pass judgment on that material. In arbitration those bottlenecks tend to be gathering evidence, collating disclosure, legal research, case preparation and drafting. Many of those tasks do not require critical thinking so much as time and resources. Arbitrators then need to digest that output and reach an informed and durable decision.
At the same time, an opportunity can be missed. In reviewing emails, letters, contracts and other documentation, reviewing witness statements, then applying that factual background to the applicable law, counsel and arbitrators build a mental image of the case beyond simply reading an AI generated summary. That absorption of the case has its benefits. This is not, or shouldn’t be, a case of an AI agent usurping the human role, like HAL in Kubrick’s 2001 A Space Odyssey refusing to open the spaceship’s outer door, saying “I can’t do that, Dave.” It is the more critical question of the bottleneck at which the lawyer or arbitrator applies the time, skills and integrity expected of a professional to each stage of the dispute.
That human involvement needs to be recognised and preserved.
AI in Arbitration
At their most basic, AI chatbots are a useful and more focused alternative to Google and other search engines, and that is perhaps the most common use. That position is changing rapidly.
AI is increasingly used throughout the arbitral process, extending from relatively straightforward document management to trial preparation with the use of Freshfield’s Virtual Witness Stand to stress-test witnesses using hostile cross-examination, identify gaps or weaknesses in testimony, and assess the resilience of competing legal theories.
At each stage, the potential for cost and time saving has become significant. In the 14th Queen Mary University of London International Arbitration Survey 2025, clients again identified cost and delay as being major concerns. In relation to AI, the surveyors concluded:
Use of AI is expected to grow significantly over the next five years, driven by the potential for efficiencies. Principal current uses of AI include factual and legal research, data analytics and document review. AI assistance in drafting and in evaluating legal arguments is also expected to increase, but significant concerns persist about accuracy, ethical issues and AI’s ability to handle complex legal reasoning.
The underlying concern, as shown in the Jevons paradox referred to above, is that in the course of preparing for a case with all its wading through documents and legal texts, lawyers build an understanding of the case and what is relevant. Similarly, for arbitrators, a considerable amount of legal analysis goes on in the background when reviewing documents and digesting legal submissions.
There is clearly a case for reducing the drudgery or repetitive activity, and largely little value added in many of those roles, particularly in the context of the increase in documentation and its complexity arising from the use of technology to date. However, in easing those boring and repetitive tasks, the opportunity for critical thinking should not be abandoned. Simply accepting an LLM’s summary of facts and legal research, setting parameters for drafting submissions without reviewing the detail, is fraught with potholes, not least AI’s tendency to generate hallucinations.
Similarly, an arbitrator asking an AI agent to summarise the documentation and counsels’ submissions falls well short of proper consideration of the case. All submissions, whether prepared with Ai input or not, should be read with care. While AI provides a perfect opportunity for increased efficiency, reduced clerical time and reduced cost, it must be to increase the quality of legal and arbitral consideration, rather than supplanting it.
At the International Council for Commercial Arbitration (ICCA) Madrid Congress last month, three core concerns about the use of AI were discussed:
(1) maintaining confidentiality;
(2) verifying all material produced using AI agents; and
(3) prohibiting the use of AI in decision-making.
Confidentiality
Where an AI agent is set a task and is:
· logged into online services (for example WestLaw, iLaw, Lexology and LexisNexis); and
· given access to the servers containing all the dispute documents (fact summaries, witness statements, documentation, submissions, expert reports and relevant documentation);
· given a task (within various parameters),
that task is still carried out online, with the risk that client information is also disclosed online. There is no guarantee that the such information, and the resulting output, will not be used again for other purposes.
The whole point of AI is its ability to learn from previous tasks. It is true that enterprise versions of most AI applications provide some comfort, but there can be no guarantee that some level of confidential information has not been released onto the internet.
Verification
The very purpose of using LLM’s is to achieve efficiency carries with it reduced qualified human input. The challenge is to preserve professional consideration where it matters. That means more than simply accepting the result, but also understanding how the AI agent got to the end result.
That is a considerable task which will need to avoid replicating the very thing the use of LLM’s was to avoid.
Decision-making
And on the last concern, we have not reached the stage where we can do away with human involvement and simply accept an AI determination of a dispute.
An arbitrator using an AI agent to summarise the submissions, rather than reading and considering them, falls well short of discharging the tribunal’s professional duty. If actually reading and considering what the parties have submitted is overwhelming, asking an AI agent to draft the determination for checking is a relatively short step. Each step reduces the quality of human input.
The central point is that the use of AI by both counsel and arbitrators provides a perfect opportunity to increase the quality of professional oversight that Tom Bingham outlined above. Protection of human involvement is not unsurmountable. It just needs to be borne front of mind when adopting what AI has to offer.
Protocols
A number of arbitral institutions are producing protocols and guidelines on the use of AI, ranging from the relatively light-handed approach of the Stockholm Chamber of Commerce (SCC Arbitration Institute Guide to the Use of Artificial Intelligence in Cases Administered under the SCC Rules 2024), which encourages the use of AI systems provided confidentiality, quality, integrity and non-delegation of the decision-making mandate are preserved, through to the Chartered Institute, with its more prescriptive guidelines (Guideline on the Use of AI in Arbitration 2025).
The CIArb AI Guideline goes into considerable detail, encouraging enquiry about the use of AI, regulation by the tribunal about its use, disclosure by both counsel and the tribunal, and recording the use of AI in the award. Much of this, while a brave attempt to regulate, goes considerably further than is necessary to simply record the duties of all involved.
At the more extreme end, the American Arbitration Association has launched its AI Arbitrator which will evaluate the merits of claims, generate explainable recommendations and prepare draft awards. That output is then reviewed, validated and revised (where necessary) before the award is issued.
As these guidelines proliferate, the core issue will be to identify and to retain the human mandate as outlined in the SCC Guide. Whether this is done in the arbitration agreement, the arbitrator’s appointment, recorded in the first procedural order or by adopting the draft clauses in the CIArb Guideline is probably of no import.
The use of AI and recognition of the importance of human input does, however, need to be recognised.
Conclusion
The future use of generative AI, and more particularly Large Language Model agents, in arbitration is inescapable. Despite its name, however, the technology does not think, nor does it understand the terms it is using; it is simply a predictive text generator, capable to carrying out numerous complex tasks more quickly, cost-effectively and potentially more accurately, than humans can. Similarly, there can be no question that clients want arbitration to be conducted more efficiently and more cost effectively. That is the core appeal of arbitration, after party autonomy.
The trick will be for counsel and arbitrators to maximise those efficiencies without surrendering the human function of critical thinking, and decision making. Provided confidentiality, accuracy of output and human oversight are safeguarded, then the use of such LLM agents will become essential.
The goal must always be to improve the quality of such oversight by reducing the laborious and high cost/low value tasks; preserving the judgment of counsel and tribunals in determining the dispute by setting the tasks to be undertaken by the LLM’s; and improving the quality of such professional assessment.
That can be achieved more by accepting the limits of what AI can offer than by setting prescriptive rules; AI will move too quickly to fill the gaps for the latter approach to be effective.
Return to All