The Other AI Problem in Arbitration: When AI Makes the Record Too Big to Read

Artificial intelligence (“AI”) can undermine arbitral decision-making in two different ways. It can do too much of the arbitrator’s work, replacing the personal reasoning the parties selected the tribunal to provide. Or it can create so much work that meaningful personal engagement with the record becomes practically impossible. The first is a delegation problem. The second is an overload problem. Both threaten the same thing: the arbitrator’s personal mandate.

The Québec Superior Court’s decision in ARIHQ v Santé Québec (“Québec”) illustrates the first risk. In April 2026, the Court set aside an arbitral award after finding the arbitrator had effectively delegated his decision-making function, relying on AI-generated material — including fabricated authorities and a mischaracterized precedent — he had not adequately verified. The decision has dominated this year’s discussion about AI in arbitration, and readers of this Blog will already be familiar with its facts (see here). But it answers only one side of the problem: what happens when AI enters the decisional core of the process. The harder question may arise earlier, when AI lets parties multiply the record the tribunal must absorb.

A Problem Already Visible Before Québec

The strain was already showing up in this Blog’s own coverage (see here) before the Québec award was even reported. This year’s Paris Arbitration Week panels flagged “the growing burden of increasingly lengthy awards,” noting it is not uncommon for decision-makers to review multiple lengthy submissions within a single week, each requiring extensive citation checks. A whitepaper discussed on this Blog (see here) in February 2026 proposed restraining document production through word or page limits on requests and objections. Both pieces were written before generative AI had fully entered the conversation as a volume multiplier rather than merely a drafting aid. That is the gap this post tries to fill.

AI as a Volume Multiplier

The point is not simply that AI helps lawyers draft faster. It lowers the cost of producing information much faster than it lowers the cost of judging information. Submissions, objections, chronologies, authorities, document requests, and responses to every point the other side makes can now be generated and repeated at a marginal cost approaching zero. AI makes advocacy scalable in a way adjudication is not: a party can respond to every argument and object to every request, while the tribunal’s attention remains human-scale.

Diego Palomo, in a note on procedural inflation, captured the underlying error: submissions do not gain value from their length, but from what they manage to communicate. Generative AI does not merely make it possible to bet on length instead of clarity. It makes that bet nearly free.

A vivid illustration of an unconstrained version of that bet did not occur in arbitration, but in a national court system, and it is worth taking seriously precisely because arbitration has fewer institutional buffers than a state judiciary. In July 2026, a lawyer in Chile used an automated tool to file 38,477 submissions in under three days, reportedly overwhelming the court system’s technical capacity. Whether the tool technically qualifies as AI remains disputed, with a local legal-tech association cautioning that automated filing does not necessarily involve it. What is not in dispute is the asymmetry it exposed: the tool made it possible to generate roughly 500 to 900 filings per court at machine speed, while the courts remained dependent on human capacity to receive and process them.

A national court system may in theory have a public budget and administrative staff built for such shocks; in practice, many do not yet have the infrastructure to process what AI-enabled advocacy can generate. Arbitration is exposed to the same pressure under a leaner institutional design: a tribunal typically has one or three arbitrators, a secretariat of finite capacity, and a cost structure — hourly, ad valorem, or otherwise — in which time spent reading the record is a real expense for the parties. If volume inflation is already testing courts, there is little reason to assume arbitration would be more resilient.

The cost of a bloated record is not measured only in pages, but cognitively: a 300-page submission requires the tribunal to change context, verify authorities, distinguish repetition from development, and remember where a supposedly dispositive point first appeared. The more AI makes production cheap, the more expensive human verification becomes, and a tribunal may appear to be receiving a fuller record when it is in fact receiving more repetition and more opportunities for peripheral material to crowd out what actually decides the case.

The Judicial Mirror

State courts are not simply absorbing this pressure passively. Germany offers a prominent example, and one already familiar to readers of this Blog from earlier coverage of the ICC YAAF workshop. The Frankfurt District Court developed FraUKe (Frankfurter Urteils-Konfigurator Elektronisch), an AI-assisted system supporting judges in mass airline-passenger litigation, where the Court receives roughly 10,000 to 15,000 such cases a year. The Stuttgart Higher Regional Court implemented OLGA (Oberlandesgerichtsassistent) to classify and cross-reference thousands of largely similar diesel-emissions appeals.

These are real, publicly documented, multi-year pilots, but they are also the exception rather than the ordinary condition of adjudication, and they point to the second half of the volume problem: once AI makes it cheap for one side to inflate the record, the institutional response may become more AI on the other side, to make it cheap to process. For a national judiciary handling thousands of near-identical claims, that trade-off may sometimes be defensible.

Arbitration cannot simply import that solution. Because the parties have selected particular arbitrators to decide their dispute, delegating the processing of the record raises a different concern than in mass state-court litigation. A growing number of arbitral institutions and professional bodies have issued guidance on this point; the SVAMC and CIArb instruments are used here only as illustrative examples. SVAMC Guidelines 6 provides that arbitrators may not delegate any part of their personal mandate or decision-making function to AI, regardless of sophistication. CIArb's Guidelines take the same direction: arbitrators may use AI to support efficiency but must maintain independent judgment and verify AI-generated information before relying on it, and tribunals may regulate parties’ AI use through procedural orders, with non-compliance carrying consequences in costs or adverse inferences.

The concern therefore extends beyond AI making the wrong decision: it includes AI determining, implicitly, which parts of an AI-expanded record the arbitrator ultimately sees. The personal mandate does not require an arbitrator to perform every administrative or information-processing task unassisted — well-established practice already permits secretarial and technological support. What it requires is that the arbitrator’s own engagement with the evidence and arguments material to the decision remain independent and substantive; the harder question is at what point assistance begins to displace that engagement. Québec shows a state court may annul an award where assistance crosses into delegation. The overload problem tests the same principle from the other side: even without delegated reasoning, AI-enabled volume may make independent engagement with the record practically impossible.

Delegation, Overload, and the Arbitrator’s Personal Mandate

Québec and the volume problem look like unrelated risks, but they are two expressions of the same concern: the erosion of the personal, considered engagement both parties are paying an arbitrator to provide. In Québec, the erosion happened on the decision side, when the arbitrator adopted reasoning it had not built or verified. On the volume side, it happens earlier and more quietly, when the sheer quantity of AI-generated material makes genuine engagement with the record practically unaffordable within the time and budget the parties agreed to.

That reframing matters for what tribunals should do about it. AI should make proportionality more important, not less. The existing toolkit already points the right way: page and word limits, limits on document requests, cost consequences for disproportionate conduct, and case-management orders directed at AI-generated volume — understood not merely as efficiency devices, but as safeguards of the tribunal’s ability to perform its personal mandate. Any such measure must also preserve each party’s reasonable opportunity to present its case; proportionality is not a license to exclude relevant material, but a discipline for keeping the record commensurate with what a fair hearing requires. The question is no longer merely whether a submission is relevant, but whether its marginal relevance justifies the cognitive cost it imposes.

Conclusion: Proportionality as Protection of Judgment

The question for arbitration is therefore no longer simply whether AI may assist the tribunal. It is also how far the tribunal may allow the parties to use AI in ways that make its own personal engagement with the case practically impossible. Québec answered what may happen when a tribunal stops doing that work itself. Arbitration must still answer what should happen when the volume the parties generate makes it impracticable for the tribunal to do that work at all. If the arbitrator's personal judgment is the product the parties have purchased, procedural proportionality may be the mechanism that protects it.