When the dispute is about AI: arbitration's new rules for AI discovery and evidence

When a dispute centers on what an artificial-intelligence system did and why, the evidence is no longer a tidy stack of emails. It is a model with millions of parameters, the data it trained on, the configuration that shaped its behavior, and outputs that may have shifted between the disputed event and the day a complaint was filed. Ordinary discovery tools strain against that target. A model is a moving, opaque, and often proprietary artifact — and the conventional answer of “produce the relevant documents” does not map cleanly onto it.

Arbitration providers reached this problem before the federal rulemakers did. In April 2024, JAMS published the first set of arbitration rules built specifically for disputes involving AI systems — the JAMS Rules Governing Disputes Involving Artificial Intelligence Systems, accompanied by a model clause and a form protective order.^[1] The rules are worth studying not because arbitration governs most AI cases (it does not), but because they are a real-world laboratory for three questions every forum will eventually have to answer: how to disclose AI use, how to handle the authenticity of AI-touched evidence, and how to keep AI discovery from swallowing the case.

The core move: take the experts to the model, not the model to the parties

The signature provision of the JAMS framework addresses the proprietary, high-stakes nature of model evidence directly. Rather than ordering a party to hand over its hardware, software, models, and training data, the rules direct that those materials be made available to one or more experts in a secured environment established by the disclosing party — and the experts may not transmit or remove any produced materials from that environment.^[2] Where the parties jointly request it, the arbitrator can appoint neutral experts from a JAMS-maintained roster, with costs ordinarily split between the parties.^[2]

This is a genuinely different posture from traditional production. It treats the model as something to be examined in place, under controlled conditions, rather than copied and circulated. That design solves several problems at once: it protects trade secrets and the integrity of the system, it keeps sensitive training data — which may itself contain personal information regulated under laws such as the GDPR or CCPA — from spreading, and it routes interpretation through people equipped to read it. A backing protective order governs the handling of everything disclosed.^[1]

Disclosure of AI use as a procedural duty

The JAMS rules answer one slice of the disclosure question — disclosure of the AI system at issue. A second, faster-moving slice is disclosure of AI use in the conduct of the proceeding itself: briefs drafted with generative tools, exhibits touched by AI, analyses produced by an algorithm. Here the AAA-ICDR's March 2025 Guidance on Arbitrators' Use of AI Tools is the cleaner reference point. It tells arbitrators to make reasonable inquiries about any AI tool before relying on it, to cross-check AI output against primary sources, and — critically — recognizes that the tribunal may require disclosure of AI use where that use could affect the proceeding.^[3]

Read together, the two instruments sketch an emerging norm: AI is not invisible infrastructure to be used silently. When it shapes the evidence or the decision, its use becomes a fact the forum is entitled to know. That norm tracks what courts are independently demanding — that a party asserting an exhibit is AI-generated, or defending one that is, put the technical record on the table rather than leaving authenticity to assertion.

Authenticity becomes a provenance question

AI-related disputes scramble the usual authenticity inquiry. The hard question is not only “is this exhibit genuine?” but “how was this output produced, and can that be reconstructed?” Models that learn continuously change their parameters and their behavior over time, so the system that generated a disputed output in March may no longer exist by discovery — a preservation and spoliation problem with no clean analog in paper litigation. Without a snapshot or an audit trail, recreating the state of the system at the moment in question can be impossible.

The secured-environment model helps here too, because it lets experts examine the architecture, configuration, and available training data rather than accept a litigant's characterization of them. This mirrors how the Federal Rules of Evidence already treat machine output: authenticity under Rule 901(a) requires evidence sufficient to support a finding that an item is what its proponent claims, and the workhorse for system output, Rule 901(b)(9), turns on showing that the process or system “produces an accurate result.”^[4] For a generative or adaptive system, satisfying that standard demands exactly the technical record the JAMS process is designed to surface. The same logic animates proposed Federal Rule of Evidence 707, published for comment in 2025, which would subject machine-generated output offered without a human expert to Rule 702-style reliability scrutiny.^[5]

Efficiency through proportionality and technical competence

AI discovery is expensive precisely because the data is voluminous, unstructured, and hard to interpret. The JAMS rules lean on two levers to contain that cost. The first is proportionality: document requests are confined to material directly relevant to the dispute, e-discovery is generally limited to ordinary business sources, and the arbitrator may deny requests where the costs and burdens are disproportionate to the nature of the dispute or the amount in controversy.^[2] That standard is a close cousin of Federal Rule of Civil Procedure 26(b)(1), which since 2015 has limited discovery to matter that is relevant and proportional to the needs of the case, weighing burden against likely benefit.^[6]

The second lever is competence at the top. Arbitration lets the parties choose a decision-maker — or a discovery referee — who already understands model architecture, training pipelines, and search-term mechanics. A neutral fluent in the technology can resolve scope, sampling, and proportionality fights faster and more accurately than a generalist forced to learn the domain mid-case, and can craft reliability rulings in an area where binding precedent is still thin. That expertise is the quiet engine behind the efficiency the rules promise.

What practitioners should take from this

Arbitration is not the destination for most AI disputes, but its rules are a preview. Three lessons travel to any forum. First, treat the model and its provenance as the evidence: preserve snapshots, configurations, and audit trails the moment a dispute is foreseeable, because an adaptive system will not wait for discovery. Second, plan for in-place, expert-mediated examination rather than wholesale production — it protects trade secrets and regulated data while still letting the truth be tested. Third, build the authentication and disclosure record early; the forum that asks “how was this produced?” rewards parties who can answer with a verifiable trail rather than an assurance. The durable advantage, as always, belongs to evidence that can prove what it is. See /provenance.