
Generative AI can undermine arbitral decision-making by making the record too large for an arbitrator to read, not only by doing the arbitrator's thinking. That is the central argument of a recent commentary by Mario Covarrubias Jurado of Estudio Garrido Abogados. He contends that procedural proportionality should be treated as a safeguard of the tribunal's personal judgment.
Two risks, one mandate
The author separates two ways AI can interfere with adjudication.
- Delegation: AI takes over reasoning that the parties entrusted to the arbitrator.
- Overload: AI generates so much material that genuine personal engagement with the record becomes practically impossible.
Both risks, he argues, threaten the arbitrator's personal mandate.
The delegation risk materialised in ARIHQ v Santé Québec. In April 2026, the Québec Superior Court set aside an arbitral award after finding that the arbitrator had effectively delegated his decision-making function. He had relied on AI-generated material that he had not adequately verified, including fabricated authorities and a mischaracterised precedent. The commentary notes that this decision addresses only what happens once AI enters the decisional core of the process. It does not address the earlier stage, when parties use AI to multiply the material the arbitral tribunal must absorb.
Cheap to produce, costly to judge
The key point is an asymmetry. AI lowers the cost of producing submissions, objections, chronologies, authorities and document requests far faster than it lowers the cost of evaluating them. A party can answer every argument and object to every request, but the tribunal's capacity to read remains human-scale.
The author cites an incident from a national court system. In July 2026, a lawyer in Chile used an automated tool to file 38,477 submissions in under three days, reportedly overwhelming the courts' technical capacity. Whether the tool qualifies as AI is disputed, but the episode produced roughly 500 to 900 filings per court. The commentary argues that arbitration has even fewer institutional buffers than a state judiciary: typically one or three arbitrators, a secretariat of finite capacity, and fee structures in which reading time is a real cost to the parties.
Some courts have answered volume with technology of their own. The Frankfurt District Court developed FraUKe to support judges in mass airline-passenger litigation, which brings roughly 10,000 to 15,000 cases a year. The Stuttgart Higher Regional Court uses OLGA to classify and cross-reference diesel-emissions appeals. The author argues that arbitration cannot simply import this approach, because the parties chose specific arbitrators to decide their dispute.
He points to soft-law instruments as illustrations:
- SVAMC Guideline 6 bars arbitrators from delegating any part of their personal mandate or decision-making function to AI.
- The CIArb Guidelines allow AI for efficiency. They require independent judgment and verification of AI-generated information, and they permit tribunals to regulate the parties' AI use through procedural orders, with possible consequences in costs or adverse inferences.
The commentary also identifies a subtler risk. AI processing tools may implicitly determine which parts of an AI-expanded record the arbitrator actually sees.
Proportionality as protection
The author's conclusion is that AI makes proportionality more important, not less. He points to familiar tools:
- page and word limits;
- limits on document requests;
- cost consequences for disproportionate conduct;
- case-management orders aimed at AI-generated volume.
He presents these as protections of the tribunal's ability to perform its mandate, not merely as efficiency devices. Any such measure must still preserve each party's reasonable opportunity to present its case. Earlier signals point the same way. Panels at this year's Paris Arbitration Week flagged the growing burden of lengthy awards, and a February 2026 whitepaper proposed word or page limits on document production requests and objections.
In the author's framing, the question is no longer only whether a submission is relevant. It is whether its marginal relevance justifies the cognitive cost it imposes on the tribunal.