The AI Arbitrator and the Two-Tier Trap: Safeguarding Legitimacy in Automated Proceedings

[Aditya is a student at Jindal Global Law School.]

On 17 September 2025, the American Arbitration Association – International Centre for Dispute Resolution (AAA-ICDR) announced an AI-assisted “arbitrator” for documents-only construction disputes, with initial deployment slated for November 2025. In this model, the system analyzes the written record and produces a draft, reasoned award, while a human arbitrator remains responsible for reviewing, modifying, and issuing the final decision. The pilot is deliberately narrow being used for only document-heavy construction claims.

This announcement has renewed familiar questions about legitimacy, party autonomy, due process, confidentiality, bias, and transparency in AI-assisted adjudication. Some of these concerns are generic to any decision support tool; others are specific to arbitration’s consent-based framework and to the enforcement standards that apply to reasoned awards. Documents-only construction cases are a logical testing ground given their volume and documentary intensity, but concentrating AI in this segment also raises distributional questions about who receives which kind of process, on what terms and with what safeguards. 

This piece addresses one central risk: the emergence of a two-tier justice system if AI-assisted workflows are applied primarily to lower-value, documents-only claims while higher-value disputes continue to receive fully human deliberation. The analysis situates the pilot within existing procedural architecture, explains how the use of AI-generated draft awards places asymmetric pressure on the core guarantees of arbitral legitimacy, and therefore proposes a governance framework to capture the speed and cost advantages without creating second-class process.

Compliance with Existing Procedure

The pilot operates inside a conventional documents-only construction arbitration. The composition of the tribunal does not change: the human arbitrator remains the tribunal for all legal purposes, and the award is issued in the arbitrator’s name. The AI functions as a drafting and analysis tool. It ingests the written record submitted by the parties, structures issues, and produces a draft set of reasons and dispositive language. That draft is then reviewed, and where necessary, revised by the arbitrator before the final, reasoned award in rendered. 

Placed against the standard architecture of international arbitration, this workflow relies on familiar building blocks. Documents-only proceedings are already a recognized format under UNCITRAL (Article 17(3)), CiARB, and ICDR (International Expedited Procedures); tribunals already exercise case-management powers to organize evidence, set timetables, and streamline drafting; and awards already reflect the tribunal’s analysis of the record and applicable law. The pilot changes how some of that synthesis is produced, not who decides or what counts as the record. These features are in consonance with AAA-ICDR Guidance on Arbitrators’ Use of AI Tools.

In procedural terms, the record remains the parties’ submissions and exhibits. The tribunal remains responsible for determining relevance, weighing evidence, and articulating reasons. The presence of an AI-generated draft does not alter the obligation to produce a reasoned award or the allocation of responsibility for that reasoning: it is the arbitrator’s reasoning that is communicated to the parties. 

Seen through the lens of supervisory and enforcement law, nothing in the pilot redefines the applicable legal regime. The lex arbitri (the law of the seat of arbitration) at the seat continues the govern the conduct and supervision of the proceedings; recognition and enforcement proceed under the usual instruments. Any later scrutiny of an award remains tied to the same questions as in a traditional documents-only arbitration: who constituted the tribunal, what the agreed procedure was, what material formed the record, and whether the final award expresses the tribunal’s reasons. 

Core Procedural Requirements – and the Two-Tier Risk

What is new here is not that a tribunal may consult software but that an “AI arbitrator” is being offered by an arbitral institution as a defined procedural feature whose primary function is to generate the first iteration of reasons and dispositive text for a human arbitrator to review and issue. By contrast, when individual arbitrators have used AI to date, it has been discretionary, assistive tooling – while authorship of the reasons remained entirely the tribunal’s and usually sat outside the four corners of the procedure. In this pilot, award-drafting by the system is designed into the workflow ex ante, described to the parties, trained on a corpus of prior construction awards, and benchmarked against human legal judgment, with the human arbitrator positioned as validator and issuer of the final award. Put differently: conventional use is an arbitrator using AI; this is an institutionally branded AI arbitrator that produces the first draft of the award. 

Although the pilot can be housed within familiar procedure, fitting inside the architecture is not the same as delivering parity of process. Once the same institution begins to resolve similar disputes through materially different modalities – one fully human-crafted, the other human-validated but machine-drafted – core guarantees of arbitration come under asymmetric pressure. 

The first pressure point is consent. It is undisputed that Arbitration’s legitimacy rests on party autonomy, but in construction contracting many relationships are governed by standard forms and unequal bargaining power. If the AI-assisted workflow arrives embedded in boilerplate or presented as the only financial sensible route, assent becomes formal rather than meaningful. The predictable effect is stratification by leverage and claim value: repeat players with options reserve fully human workflows for higher-stake cases, while smaller counterparties are channeled into the AI track. The law may regard both paths as “arbitration”, but the allocation of process begins to track bargaining power rather than legal need. 

The second pressure point concerns the right to be heard and the duty to give reasons. Documents-only cases already compress participation; inserting a model that drafts the first iteration of reasons increases the distance between parties and the adjudicative mind. In arbitration, accuracy alone is not the currency of legitimacy. Parties comply because they believe they were heard by the decision-maker and can trace how the evidence produced the outcome. Where one cohort receives reasons composed directly by a human tribunal and another receives reasons composed by a system and then validated by a human, the institution delivers different experiences of adjudication. That divergence matters at two junctures. It matters at the voluntary-compliance stage, because perceived thinness of engagement erodes acceptance even when the merits are sound. And it matters later if enforcement is contested, and a party claims it lacked a meaningful opportunity to present its case or to understand how dispositive inferences were drawn from the record. 

The third pressure point arises from the date inheritance of any system trained on past awards and iteratively tuned with expert input. Stabilizing doctrine can be a virtue, but it can also ossify patterns that historically favored institutional regulars. Repeat players learn which arguments the system tends to find persuasive and optimize their filings accordingly; one shot parties do not. Concentrating AI first and most intensively on lower-value claims amplifies the asymmetry precisely in the segment least buffered by extensive advocacy or generous tribunal time. Relatedly, AI-mediated workflows carry distinct security and integrity risks: inadvertent reliance on extra-record material, prompt-injection through party submissions, and other forms of adversarial manipulation. There are manageable in principle, but they are least likely to be managed with maximal conservatism on fast tracks, further widening the gap between tiers in the trustworthiness of process. 

Finally, there are enforcement optics. Segregation by value invites different litigation postures. Higher-value, human-led awards are less likely to attract process objections; lower-value, AI-assisted awards are more likely to face targeted challenges framed in terms of composition or procedure “not as agreed,” inability to present one’s case, or public-policy concerns about opacity and extra-record contamination. Even if such challenges ultimately fail, they impose delay and cost that dilute the very access-to-justice gains the pilot is designed to secure. In short, the pilot’s placement within documents-only, lower-value construction claims supplies precisely the structure conditions under which consent is thinner, reasoning is less elaborated, and risk-bias security, and challenge – is more concentrated. That is the essence of a two-tier regime. 

Governance Framework 

The way to avoid that stratification is not to abandon the pilot, but to specify it. A legitimate AI-assisted workflow needs to be framed in procedural terms that make the human arbitrator’s authorship, the parties’ hearing rights, and the evidentiary boundary unmistakable. 

The starting point is consent: the use of an AI-drafted first pass should be confirmed after the dispute arises, in writing so that agreement reflects real choice rather than boilerplate inertia. That same procedural order should describe, in plain terms, what the system does and what it does not do. Because legitimacy in arbitration is carried by reasons, parties should receive a reasoned award that makes clear it is the tribunal’s reasoning, and they should be able to see how the tribunal engaged with the draft – an appendix that cites the record relied upon and indicates material changes made by the arbitrator closes that loop. 

Disclosure and control are the next pillars. An institutional tool that drafts reasons is not the arbitrator’s private tool; it is part of the procedure. The model family and version, the fact of benchmarking against prior awards, and the handling of party data should be disclosed at the outset, with the model version constant for the life of the case so that the decisional context does not shift mid-stream. A basic audit trial – what was ingested, what the system produced, what the arbitrator altered – preserves reviewability without turning the process into discovery on the model’s internals. 

Fairness also requires attention to distributional effects. Before deployment and at regular intervals, the institution should test for outcome drift and repeat-player advantages across comparable claims and publish metrics that matter to users: cycle time and cost per claim, dispersion of outcomes for similar fact patters, and the rate and locus of human overrides. Where the system signals low confidence or where a party credibly points to an omission, the case has to default to human-first analysis rather than an AI-drafted path. 

Finally, confidentiality and integrity cannot be assumed. The workflow must operate in a secured environment that prevents extra-record contamination, resists adversarial inputs in party filings, and segregates confidential annexes where appropriate. Taken together – real consent, transparent authorship, version control and logs, fairness testing with publication of core metrics, triggers for human-first handling, and hardened data practices – these measures retain the speed, and consistency gains of the pilot while foreclosing the emergence of a second-class tier of justice. 

Conclusion

As the CEO and President of the AAA-ICDR, Ms Mary McCormack, rightly put, “innovation is no longer optional – it is essential.” An AI-assisted track can indeed make arbitration faster and more accessible. Whether it also strengthens legitimacy, however, will depend on its execution. If the rollout keeps faith with the essentials of a legitimate arbitration, it will set a credible example for expansion beyond construction. If it does not, the field will inherit speed at the price of uneven justice. Execution and not aspiration will determine whether this innovation becomes an equalizer or a fault line in modern arbitration.