The AAA's Own AI Arbitration Survey Puts Practitioner Trust at 2.18 Out of 5

The AAA now has numbers on how its own neutrals use AI, and the headline number is 2.18 out of 5

The American Arbitration Association (AAA) and the legal research company Jus Mundi released The State of AI in U.S. Arbitration 2026 on 23 September 2026, a survey of 557 US arbitration professionals that puts average trust in AI at 2.18 out of 5.00 — and at 3.42 among daily users against 0.83 among those who never touch it. The practical significance of this AAA AI arbitration survey is not the aggregate score but the spread: on the provider's own data, whether a neutral or an advocate regards AI as usable turns almost entirely on whether they have used it, and the population of independent arbitrators that counsel actually select from is the least governed group in the sample, with 15% reporting a clear AI policy against 83% at global law firms.

The report arrived five days after the AAA seated 37 lawyers in an AI Ambassador Program organized into four working groups, and one week before Governor Gavin Newsom's 30 September 2026 deadline to act on California SB 574, which would bar an arbitrator from delegating any part of the decisionmaking process to a generative AI tool. Within a fortnight, the largest US provider has produced the empirical baseline, the institutional working machinery, and — if the bill is signed — the first state statute telling arbitrators how to behave. For anyone drafting an arbitration clause, choosing a neutral, or weighing a challenge to an arbitral award, the three documents need to be read together.

What did the AAA and Jus Mundi AI arbitration survey find?

The survey reached 557 US arbitration professionals — arbitrators, legal practitioners and in-house counsel — and is presented as the first comprehensive, data-driven benchmark of AI adoption across the US arbitration community. Its central claim is that experience, not seniority or role, is what moves attitudes: professionals who use AI regularly report materially higher trust than those who do not, while the concerns of experienced users become narrower and more specific rather than disappearing.

Five figures carry the report.

Finding	Figure	What it measures
Average trust in AI, all respondents	2.18 / 5.00	Below the midpoint of the scale
Average trust, daily users	3.42 / 5.00	Above the midpoint
Average trust, non-users	0.83 / 5.00	Near the floor of the scale
Respondents naming inaccuracies and hallucinations as a concern	78%	Holds across experience levels
Respondents with a clear organizational AI policy	83% at global firms; 15% of independent arbitrators	Governance coverage

Two further findings bear on how the profession will spend the next year. Reported efficiency gains run inversely to firm size: 66% of respondents at boutique firms reported significant time savings, against 51% at mid-size firms and 50% at global firms. And when respondents were asked what would accelerate responsible adoption, the top two answers were institutional guidelines and best practices (30%) and arbitration-specific AI tools (24%) — in other words, the profession is asking providers to do the work, and it asked the provider that commissioned the survey.

Respondents also expect AI to absorb routine work while raising, not lowering, the premium on human judgment, expertise and advocacy. That is a comfortable conclusion for a provider to publish, but it is consistent with where the reported use cases sit: research, analysis and document handling rather than adjudication.

Do arbitrators trust AI? The answer is an experience gradient, not a number

On this data the honest answer is that arbitration professionals as a class do not trust AI — 2.18 out of 5.00 is a vote of limited confidence — but the class average conceals a gradient steep enough to make the average misleading. Daily users sit at 3.42; non-users sit at 0.83. The gap of roughly 2.6 points is larger than the distance between the aggregate score and either pole.

That matters for tribunal composition. A three-member panel drawn from this population can easily contain one neutral who drafts with an AI assistant every day and one who regards the tools as unusable, with no shared baseline about what either may do with them. Nothing in the standard commercial rulebooks resolves that disagreement between co-arbitrators; it gets resolved, if at all, by a chair's procedural order or by the parties raising it in the first case management conference.

It also cuts against a common assumption in the AI-and-arbitration literature — that resistance reflects unfamiliarity and will decay on its own. The report's own framing is subtler: as practitioners gain experience, their understanding of the risks becomes more defined and focuses on problems encountered in practice. Experience does not dissolve the objections; it relocates them from the abstract to the operational.

Where the governance gap sits: 83% of global firms, 15% of independent arbitrators

The single most actionable number in the report is the policy-coverage split. Institutional respondents are largely covered: 83% at global firms said their organization has a clear AI policy. Independent arbitrators — the sole practitioners, retired judges and boutique neutrals who make up much of any provider's roster — reported 15%.

The asymmetry is structural rather than culpable. A global firm has a general counsel's office, a risk committee and a procurement function; an independent neutral has none of those and no client pressing a policy on them. But the consequence lands on counsel. In a bilateral commercial arbitration the party that wants to know whether the tribunal is running submissions through a general-purpose model has only one reliable mechanism: asking, on the record, early, and getting the answer into a procedural order.

Where confidentiality obligations are contractual, the gap is sharper still. A Confidentiality Provision in the arbitration agreement, or a protective order covering trade secrets, does not distinguish between a paralegal and a model; uploading the other side's confidential submission to a consumer tool is a disclosure. The 15% figure is a measure of how many neutrals have written that rule down.

What is the AAA AI Ambassador Program?

The AI Ambassador Program, announced on 18 September 2026, brings together 37 attorneys from leading law firms, organized into four working groups, to identify emerging legal and procedural issues arising from AI and to produce practical resources — articles, white papers, webinars and other educational material — for users, counsel, arbitrators, mediators and other professionals. The four workstreams are AI evidence and arbitration procedure; emerging AI disputes; AI-driven commerce and automated transactions; and digital assets and algorithmic finance. Announced by AAA President and Chief Executive Bridget M. McCormack, the program's stated premise is that AI is changing both the kinds of matters entering arbitration and the evidence, procedures and decision-making tools used to resolve them.

Read against the survey, the Ambassador Program is the provider's answer to the 30% who asked for institutional guidelines. Two cautions are worth recording. First, the program's announced output is educational, not regulatory: white papers and webinars are not rules, and nothing in the announcement commits the AAA to amending its commercial or consumer rulebooks. Second, the first working group's subject — the authenticity of AI-related evidence — is the one that will reach hearings soonest. Deepfaked audio, synthesized documents and model-generated exhibits arrive in arbitration with no equivalent of Federal Rule of Evidence 901's accumulated case law, and with Discovery in Arbitration narrower than in court, a party's ability to test provenance is correspondingly thinner.

Are arbitrators allowed to use AI? The provider instruments, in order

Yes, subject to conditions the providers have written down over the past eighteen months rather than any statute now in force. The AAA-ICDR issued its Guidance on Arbitrators' Use of AI Tools in March 2025, setting out four core principles: ensuring the accuracy and reliability of AI-generated information; maintaining fairness and Due Process (Arbitral); preserving independent decision-making; and being transparent with the parties. The operative requirement is that arbitrators retain complete control over decisionmaking and cross-reference outputs against primary sources.

Instrument	Date	What it does
JAMS AI Disputes Clause, Rules and Protective Order	Effective 14 June 2024	Opt-in rules for disputes about AI systems; AI systems and related material, including models and training data, disclosed to mutually agreed experts rather than the opposing party
AAA-ICDR Guidance on Arbitrators' Use of AI Tools	March 2025	Four principles: accuracy and reliability, fairness and due process, independent decision-making, transparency
AAA-ICDR AI Arbitrator	Announced 17 September 2025; released November 2025	Opt-in, documents-only construction disputes, human-in-the-loop; early testing reported 20–25% faster resolution and cost savings of 35% or more, with 2026 expansion to further subject matter and higher amounts in dispute
AAA Resolution Simulator	March 2026	Single-party, documents-only commercial and construction disputes; an AI-generated simulated decision using the AI Arbitrator's reasoning, for information only
AAA AI Ambassador Program	18 September 2026	37 attorneys, four working groups, educational deliverables
California SB 574 (new Code Civ. Proc. § 1282.1)	Passed 31 August 2026; gubernatorial deadline 30 September 2026	Would bar delegation of any part of arbitral decisionmaking to generative AI, require disclosure of off-record AI material, and place full responsibility for the award on the arbitrator

The distinction the AAA is drawing runs between AI as an assistive tool for a human neutral, which its guidance permits under conditions, and AI as the adjudicator, which it has productized separately and only where both parties opt in on a documents-only record. Those are different consent questions, and a clause that simply incorporates provider rules by reference — Incorporation of Provider Rules — does not answer either of them.

Can an arbitration award be vacated because the arbitrator used AI?

Not on the fact of AI use alone. Vacatur under the Federal Arbitration Act (FAA) runs through the closed list in 9 U.S.C. § 10(a), and there is no AI ground. A challenger has to map the conduct onto an existing one, and two are plausible. Under § 10(a)(4), an award may be vacated where the arbitrators exceeded their powers — the natural home for an argument that the neutral delegated the adjudicative function itself, because the parties bargained for the judgment of the person they appointed. Under § 10(a)(3), an award may be vacated for misbehavior by which a party's rights were prejudiced — the natural home for an argument that the tribunal decided on off-record material the parties never saw and could not answer.

Both routes are hard, and deliberately so. Manifest disregard survives, if at all, in narrow and circuit-dependent form, and a party will rarely have evidence of what a neutral typed into which tool. That evidentiary problem is precisely why the transparency limb of the AAA-ICDR guidance and the disclosure limb of SB 574's proposed section 1282.1 matter more than their enforcement mechanisms: they convert an unknowable into a record. A neutral who discloses AI use, and the parties' opportunity to comment, has largely inoculated the Arbitral Award. A neutral who does not has left a challenger the argument that there is nothing in the record to rebut.

For cross-border matters the same logic carries into enforcement, where a resisting party invokes the New York Convention's Article V grounds — inability to present one's case, or a procedure departing from the parties' agreement — rather than the FAA.

What the survey means for mass arbitration

Mass Arbitration is where the survey's numbers meet the largest volumes, and the report does not address it. The connection is direct. AAA and JAMS both responded to mass filings by putting a gatekeeping obligation on claimants' counsel: under the AAA's mass arbitration framework, a separate Demand for Arbitration must be filed for each claimant, accompanied by an Affirmation Requirement — a sworn declaration from counsel that the information in the demand is true and correct to the best of the representative's knowledge. The point of that affirmation is Claimant Vetting: it gives respondents a mechanism to challenge demands filed for people who do not exist or have no claim, and it gives the Process Arbitrator something concrete to rule on.

An affirmation is a personal attestation by a human lawyer. Where intake, eligibility screening and demand drafting are automated at the scale mass filings require — thousands of demands assembled from claim forms — the 78% of the profession worried about inaccuracies and hallucinations are worried about exactly the failure mode that produces an unsupportable affirmation. A batch of AI-assembled demands with a signed declaration on top is not a technology problem; it is a Rule 11-adjacent exposure for the signing firm and an obvious first challenge for respondent's counsel before the process arbitrator.

The same dynamic runs on the respondent side. Batching protocols and staggered processing exist to manage volume; if a respondent's screening of thousands of demands is model-driven, an erroneous mass rejection creates fee-non-payment and administrative-closure risk of its own. On this data neither side of the mass arbitration bar can assume the other's pipeline is governed: 15% policy coverage among independent neutrals is a low number, and the report gives no reason to think intake vendors are better.

The 2024 baseline and the hallucination backdrop

The 2026 figures only read as movement against what came before. A survey of members of the National Academy of Arbitrators conducted in the fall and winter of 2024 by Professors Harry C. Katz of Cornell and Mark Gough of Penn State — roughly 600 members polled, 219 responses, published in February 2025 — found that 87% of respondents reported not using AI in their neutral work at all. That population is labor arbitrators rather than commercial neutrals, so the two samples are not interchangeable. But a profession that was 87% abstaining in late 2024 and now returns a daily-user cohort trusting AI at 3.42 out of 5.00 has moved fast, and providers are building for the movement rather than the average.

Internationally the expectation gap is wider still. The 2025 International Arbitration Survey by Queen Mary University of London and White & Case, with 2,402 respondents, found 90% expecting to use AI for research, data analytics and document review, 54% naming time saving as the biggest driver, 51% naming the risk of AI errors and bias as the main obstacle, and 52% predicting that arbitrators will increasingly rely on AI.

The counterweight is the sanctions record accumulating in the courts. The public database of AI hallucination cases maintained by Damien Charlotin recorded 1,598 decisions involving AI-fabricated citations or content as of 9 June 2026, against roughly 200 a year earlier, with US courts imposing more than $145,000 in AI-filing penalties in the first quarter of 2026 alone and a record single-matter penalty of about $109,700. Arbitration produces no comparable public record, because awards are not published and Confidentiality Provisions keep the failures private. The absence of an arbitral hallucination docket is an artifact of confidentiality, not evidence that the problem stops at the courthouse door.

What it means for drafters, respondents' counsel, claimants' firms and neutrals

For drafters. The clause is now the cheapest place to settle AI questions. A clause that incorporates provider rules without more leaves open whether a documents-only AI adjudication is available, whether the tribunal may use assistive tools, and what must be disclosed. Parties that care should say so: an express statement that the award must be rendered by the human arbitrator appointed, and an express carve-out or opt-in for institutional AI adjudication, are both short additions. Contracts governed by California law need to be read against section 1282.1 if SB 574 is signed.

For respondents' counsel. Two additions to the first case management conference checklist: a question on the record about the tribunal's AI practices and any organizational policy, and a proposed procedural order covering disclosure of AI use, treatment of confidential material and authentication of exhibits whose provenance is contested. Doing it at the outset costs nothing; doing it after a bad award is a vacatur argument built on nothing.

For claimants' firms running mass filings. The affirmation is the exposure. Any automation in intake, eligibility screening or demand assembly needs a documented human verification step before signature, and a retained record of that step, because the process arbitrator's first question in a contested mass filing will be how the affirmation was formed.

For neutrals. The 15% figure is an invitation. A short written AI protocol — what tools are used, what never touches a confidential submission, what gets disclosed to the parties — costs an afternoon and answers the question counsel are about to start asking. It also tracks the four AAA-ICDR principles, which is the standard a reviewing court is most likely to be shown.

For institutional users and funders. The AAA's own governance benchmark, From Principles to Practice, published on 14 May 2026 from 500 senior legal and executive leaders at large US and Canadian organizations — 70% with revenue of $1 billion or more — reported the same pattern one level up: policies written, ownership assigned, and a gap between governance on paper and governance in practice. The arbitration survey is that finding applied to the people who decide the cases.

FAQ

What did the AAA and Jus Mundi AI arbitration survey find?

It found average trust in AI of 2.18 out of 5.00 among 557 US arbitration professionals, rising to 3.42 among daily users and falling to 0.83 among non-users, with 78% naming inaccuracies and hallucinations as a concern and organizational AI policies reported by 83% of global-firm respondents but only 15% of independent arbitrators.

How many people were surveyed, and who were they?

557 US arbitration professionals, described as arbitrators, legal practitioners and in-house counsel. The report was published on 23 September 2026 and is presented as the first data-driven benchmark of AI adoption across the US arbitration community.

What is the AAA AI Ambassador Program?

A program announced on 18 September 2026 in which 37 attorneys from leading law firms, organized into four working groups, will identify emerging legal and procedural issues arising from AI and produce articles, white papers, webinars and other practical resources. The four subjects are AI evidence and arbitration procedure, emerging AI disputes, AI-driven commerce and automated transactions, and digital assets and algorithmic finance.

Are arbitrators allowed to use AI?

Yes, under conditions set by provider guidance rather than statute. The AAA-ICDR's March 2025 Guidance on Arbitrators' Use of AI Tools permits use subject to four principles — accuracy and reliability, fairness and due process, preserving independent decision-making, and transparency with the parties — and requires the arbitrator to retain complete control over decisionmaking.

Can an arbitration award be vacated because the arbitrator used AI?

Not for AI use as such. A challenger must fit the conduct into 9 U.S.C. § 10(a), most plausibly § 10(a)(4) if the arbitrator delegated the decision itself or § 10(a)(3) if the award rests on off-record material the parties could not address. Disclosure at the time largely forecloses both arguments; silence preserves them.

Does the AAA use AI to decide cases?

Only where parties opt in. The AAA-ICDR AI Arbitrator, announced on 17 September 2025 and released that November, handles documents-only construction disputes with human-in-the-loop oversight, with early testing reported at 20–25% faster resolution and cost savings of 35% or more and expansion planned in 2026. The Resolution Simulator, added in March 2026, produces an informational simulated decision for a single party and decides nothing.

What does the survey mean for mass arbitration?

It supplies the risk baseline for automated claim intake. The AAA's mass arbitration framework requires a sworn affirmation from claimants' counsel for each demand, and the 78% of professionals worried about inaccuracy are worried about the failure mode that makes such an affirmation unsupportable — which is why any AI-assisted intake pipeline needs a documented human verification step before signature.