THE AUTHOR:
Jack Levin, Principal at Levin ADR
I want to use artificial intelligence (“AI”) in my arbitration and mediation practice. I also want counsel to trust me when I do. Right now, I cannot have both.
The conversation among arbitrators and mediators about AI is usually framed as a question of whether you are for it or against it. That misses the point. For a neutral, the real questions are practical. What technology is available to us? What technology are we permitted to use? Once those are answered, what do we use it for, and what do we trust it for?
Those last two questions are not the same. I can use a tool for a task without trusting it to get the answer right, as long as I am in a position to check its work. The line between using and trusting is where the profession’s thinking needs to be sharpened.
Convenience Versus Contribution
Uses of AI divide into two categories.
The first is convenience: tasks I could do myself, or that an assistant could do for me. Building a chronology is the obvious example. Pulling cases is closer to the line, because deciding what a case means is a judgment call. But checking whether a tool found the right cases is not itself an exercise of adjudicative judgment. I can verify the output on convenience tasks. I rate AI highly here, but my trust is conditional on checking.
The second category is more interesting. I call it contribution. A contribution adds to an arbitrator’s core function, which is to deliberate and decide. On contribution tasks, I do not need to trust the tool, because it never supplies the answer. It supplies questions, and I answer them. The tool may prompt, but it may not conclude.
Here is an illustration from my personal life, which is where I use AI now. I use a general-purpose AI model constantly on political and geopolitical questions. I ask it things like: Has anyone made this argument before? What am I missing? Here is a quote; is there something else I should consider? It does not do my thinking for me. It may confirm, add to, or challenge my thinking. I don’t need to trust it for that.
If I could, I would ask that kind of question of a model that holds a complete, secure case record: “Are there other places in the record that might bear on the fraud claim?” That would be enormously valuable. I already know the record. The tool points to pages; I read them and decide what they mean. If it points to nothing useful, I have lost a few minutes. If it points to something I missed, my work is enhanced.
The tool is a check on my mastery of the record, not a substitute for it.
Editing is another contribution. If I draft an award and ask a tool to comment on it, the useful response is not merely “make this shorter”. It may flag a gap in the reasoning, an inconsistency between findings, or a point I failed to address. Each of those is a question I must answer myself. The award remains mine.
So is structure. However complex the case, I would have no hesitation asking a tool to propose an outline for an award. It may offer something obvious or something outlandish. Because I know the record, I can tell which is which, and the decision about what goes where is still mine.
Arbitrators Work in Solitude
Unlike almost anyone else in the legal business, arbitrators and mediators work alone. Many judges have law clerks. Most arbitrators do not. Independence is part of the job, but solitude rarely helps. A lawyer can walk down the hall and test an idea on a colleague. A sole arbitrator usually cannot. That solitude creates the risk of error, or simply of less than the best work. An award can survive a vacatur motion and still be a source of embarrassment if the arbitrator missed something a good assistant would have caught. Approved, secure AI tools could offer real relief from that isolation. But while AI is sweeping the practice of law, considerations of security, confidentiality, and cost are leaving arbitrators with little in the way of approved AI resources.
An effective arbitrator should be a dispute resolver, not just a stone wall that listens and answers only the questions put to it. Counsel want to know what the tribunal is thinking. I will ask counsel: Are these the issues? Are these your arguments? Sometimes counsel correct me.
Sometimes a face lights up, or falls. A well-managed arbitration can help counsel and client see the dispute in a way they could not on their own. The better and the earlier the arbitrator knows the record, the better the arbitrator can manage the process. I believe arbitrations are opportunities for learning, not just for obtaining decisions. Counsel and parties who get a clear view of their situation are better able to consider settlement along the way.
Naturally, this is even more true in mediation. Deep command of a large record is usually a luxury for a mediator. If I could ask questions of a full record securely, I could have an interactive, working knowledge of it from day one. That might help me delve into hard issues and find common ground between the parties sooner.
What I Will Not Delegate
A colleague of mine owns a car that drives itself. We have had lively conversations about whether he trusts it for a three-mile or a thirty-mile trip. Driving requires judgment, yet we are moving toward reliance on AI to do it. Arbitrators should not follow that path. We should never delegate the thinking about the problem, even though the tools now make that possible.
The value parties receive from an arbitrator lies in the struggle. How many of us have sat down to draft an award convinced on Monday that we would decide it one way, gone back to the record on Tuesday, and by Wednesday felt genuinely uncertain? With time and experience, we resolve it, and the result is better for the struggle. There is no shortcut to an award.
When I say an award is mine, it has to originate with me. It cannot arrive from somewhere else and simply be edited. If a tool analyzes the facts or the law and I adopt its conclusions, I am not giving the parties the value they hired me for.
For the same reason, I am skeptical of AI summaries of critical documents, even when an institution provides them. This may seem inconsistent with my enthusiasm for querying the record, but the two are different in kind. A query supplements reading I have already done. If it misses something, I still have my own knowledge and the record itself. A summary replaces the reading, and its omissions are invisible. Why should I rely on a machine’s summary of a statement of claim or the contract at the center of a dispute? I should read the source. If summaries are useful, a better approach may be to ask the parties to prepare summaries of their own key documents. At least then we know they were written by people who are familiar with the material and are obliged not to misrepresent it.
There are products on the market that load a judge’s prior opinions, the pleadings, the relevant case law, and the case record into a closed model, and then draft a ruling in that judge’s own voice and reasoning. There is a legitimate efficiency in referring back to analyses the judge has already worked through in similar cases, but the judge still needs to bear down on the case at hand. If a developer tells me a machine can do the struggling for me, my answer is: show me. I would not rule it out forever. Perhaps one day.
Every tool is just a tool until you abdicate to it. Then it becomes something else.
The Permission Problem
For all the talk about getting to know and using AI, as far as I know, I am explicitly permitted to use only one tool in my neutral work, a product approved by a major arbitration provider. I am discouraged from using any of the general-purpose models on a case. So I do not. I understand that I should not use the consumer versions of these tools, which are not built to the confidentiality and security standards our cases require. The enterprise versions can be prohibitively expensive for an individual arbitrator.
Some arbitral institutions and professional bodies have published thoughtful principles on AI. But principles are not permissions. They tell me to protect confidentiality, disclose appropriately, and not delegate my decision. They do not tell me whether I may put a particular record into a particular product. Providers seem reluctant to tell arbitrators what we cannot do, and equally reluctant to tell us what we can.
Neutrals are also at a structural disadvantage. Counsel can clear a tool with their own clients, and a firm can spread the cost of an enterprise license across many lawyers and matters. A neutral must satisfy both disputing sides at once that their and their clients’ interests in security and confidentiality are protected, but the neutral usually bears the cost alone.
Discussing Risks and Rewards
Alternative dispute resolution (“ADR”) and AI are a natural fit. The institutions can do a tremendous service to their customers, to neutrals, and to the cause of alternative dispute resolution by facilitating conversations among all concerned to improve access to and productive use of AI.
There are two complementary routes to getting approved tools into neutrals’ hands. The first runs through the providers, and I return to it below. The second runs through counsel. Counsel already have reliable tools, and they may be able to offer the neutral access to them safely and securely. Lawyers, after all, are the demand side for AI tools. They can talk to AI tool providers about customizing products for use by neutrals.
That second route – through counsel — needs one safeguard. The questions a tribunal asks of the record reveal how it is thinking. If one side’s firm hosts the tool, that firm’s system holds the tribunal’s queries, and the other side may reasonably worry about both access and appearances. So any such arrangement should rest on both sides’ consent, with the neutral’s work walled off from the offering firm and the arrangement fully disclosed. A jointly selected or provider-hosted tool avoids the problem altogether. Either way, the tool, the use, the cost sharing, and the consent can all be built into the procedural order at the beginning of the matter.
Party autonomy is arbitration’s great strength. We are dealing with consenting adults, and if they agree to a procedure, the neutral can largely stop worrying, unless the procedure is unethical.
Perspective matters here. We lived without AI until now. We don’t need AI to do our work, but we should not let old ways of thinking deprive everyone concerned of its benefits.
AI is already being tested in small, simpler matters to generate awards under the supervision of a human arbitrator. The American Arbitration Association has innovated with its “AI Arbitrator,” an opt-in tool that until recently used AI to streamline two-party, documents-only construction disputes while keeping a human expert in charge of the final decision. The AI Arbitrator is now available for eligible two-party, documents-only business-to-business commercial disputes across industries. We may reach the point where parties to even more complex, higher-stakes matters agree to a faster, cheaper, AI-drafted award reviewed and signed by a human arbitrator, bargaining for a result rather than for reasoning. After all, arbitration is a creature of contract, so who is to say what parties can and cannot bargain for when it comes to AI? If parties want that, it is their bargain to make. It is not the service I offer.
There are unanswered questions. An arbitrator’s duty to decide is a professional obligation, supported by experience and bound by ethics. Courts asked to vacate or confirm awards will eventually draw those lines.
The Ask
Give neutrals AI tools that counsel trust, so that we can do our jobs more efficiently and effectively. Concretely, I would ask providers to do three things.
First, publish and update a list of tools that have been vetted for security and confidentiality, and say plainly that neutrals may use them.
Second, add a standard AI provision to the preliminary hearing checklist, covering the tool, the scope of use, cost, and consent so the conversation happens at the start of every case rather than never.
Third, use the size and scale of providers to negotiate enterprise access for the neutrals on their panels. That addresses the economic disadvantage directly.
No one has more contact with the users of arbitration and mediation services than providers. The better they understand what users need, the better off everyone in the process will be.
Ease of use matters, too. As with all new technologies, some AI tools require real homework to learn, and busy practitioners do not want to do it. We know that if we wait, the technology will improve and become easier to use. Secure, institutionally approved, easy-to-use tools will win over colleagues who say they have never used AI and never will.
There is nothing to fear from AI that is secure, approved, and kept in its place. For those of us who work with counsel and parties in a culture of party autonomy, it is a way for everyone to be nimbler and more productive. Disputes will not go away. We will be better prepared to resolve them.




