AI in ADR: The eDiscovery Neutral's Toolkit

Mediation and arbitration were once prized for their light procedural footprint. Parties chose them, in part, to avoid the grinding document exchange that defines federal litigation. That bargain has quietly eroded. The same electronically stored information that drives a court case now drives the private one: email, chat logs, database exports, cloud backups, system telemetry. In a data-breach dispute, the contested facts live almost entirely in machine-generated records, and no party will resolve on faith that the other searched honestly.

So ADR now carries litigation-grade discovery expectations without litigation's standing rules to govern them. There is no presiding district judge to set protocol, no local rule to default to, and often no shared definition of what a 'reasonable' search even means. The flexibility that made ADR attractive becomes the thing parties contest: scope, custodians, search terms, and the burden of production.

Into that gap steps the neutral. In a breach matter, the mediator or arbitrator is frequently the only person positioned to set a discovery protocol both sides will accept as fair before either has seen the other's data. That is a technical task as much as a procedural one, and increasingly it is a task of supervising how AI is used.

The workhorse of modern review is technology-assisted review, often called predictive coding. Instead of having lawyers read every document, a modest set of human coding decisions trains a model, and the model extends those judgments across millions of records, ranking each by likely relevance. Newer generative tools go further, summarizing document families, grouping themes, and drafting issue chronologies. In a breach case, that might mean separating routine IT tickets from the handful of messages where an engineer first flagged anomalous traffic.

Used well, these tools compress months of review into days and cut cost sharply, which is exactly why they suit ADR's promise of an expedited, proportionate resolution. But the efficiency carries consequence. A TAR system is only as sound as its training set, its seed selection, and the relevance standard the reviewers applied. Each is a human choice the machine then amplifies at scale.

That amplification is the point a neutral must hold onto. When a model propagates a coding decision across a corpus, it propagates the error along with the insight. A skewed seed set, an under-inclusive search string, or a relevance call that quietly excludes a category of records does not announce itself in the output. It simply yields a smaller, cleaner-looking production that happens to omit the documents the other side most needs.

The neutral's central function in a breach dispute is not to operate the AI or to second-guess every relevance call. It is to keep the technical record honest: the documented chain of choices behind each production should be visible, reasoned, and capable of being tested. A production earns trust not because it is large but because the method behind it can be explained.

In practice, the neutral can insist on a protocol that captures inputs, not only outputs: which custodians and data sources were in scope, how the seed set was built, what relevance criteria the reviewers used, what sampling or validation confirmed recall, and where AI was used at all. None of this requires disclosing privileged strategy. It requires disclosing method, much as an expert discloses the basis for an opinion without surrendering the client's confidences.

When the method is on the record, disputes shrink. A party arguing the other side's search was under-inclusive can point to a specific term or a missing data source rather than trade accusations of bad faith. The neutral can then make a narrow, reasoned ruling instead of refereeing an unfalsifiable quarrel. It is the instinct a court-appointed eDiscovery special master brings to a federal case, carried into the private, consensual frame of ADR.

The two recurring failure modes of AI-assisted review map onto the two arguments parties raise. The first is bias: if the training inputs lean one way, the model inherits that lean, and the production reflects it. The second is under-inclusivity: a defensible-looking set quietly missing a slice of relevant material because the selection criteria never reached it. Both are invisible in the deliverable and detectable only in the method.

This is why disclosure of AI use, not merely AI accuracy, is the operative issue in ADR. These proceedings run on the parties' confidence in one another and in the neutral. When one side has used a model to drive review or a generative tool to summarize evidence, the other side and the neutral should know that it was used and where. Silence converts a routine efficiency into a fairness objection waiting to surface, often at the least convenient moment.

JAMS has issued rules and guidance addressing AI in arbitration, and the bar's duty of competence already obligates counsel to understand the benefits and risks of the technology they deploy. A neutral need not wait for a party to raise the point. Setting an early expectation that material AI use will be disclosed, and that the method behind a production is open to examination, forestalls the harder confrontation later.

Breach disputes are uniquely awkward because the evidence is often the very sensitive data whose exposure created the dispute. Producing it for review, loading it into a platform, and feeding it to AI tools each create fresh handling risk. Cross-border matters compound the problem, where GDPR, the CCPA, and similar regimes impose duties that do not yield to the parties' desire for speed.

The neutral is well placed to insist that the review environment itself belong in the protocol: where the data resides, who may access it, whether AI tooling transmits content to external services, and how the production set is secured and ultimately disposed of. In a matter that exists because data was mishandled, a discovery process that mishandles it again is not a neutral posture; it is a second incident.

Treating the platform and its AI features as part of the record, rather than as invisible plumbing, keeps the resolution from compounding the harm it was convened to resolve.

A neutral who wants to keep AI-assisted review fair can do most of the work at the outset, in the protocol, rather than later in motion practice. Establish early that data sources and custodians will be identified, that use of TAR or generative tools will be disclosed, and that the validation behind a production can be examined on a reasoned showing. Reserve the neutral's own technical questions for method, not strategy.

Pair that with proportionality. ADR's value is a resolution sized to the dispute, and the neutral's leverage to curb an overbroad demand is strongest when the search method is transparent enough that 'proportionate' can be measured against something concrete. Drawing on forensic and eDiscovery specialists, retained by the parties or as the neutral's own resource, turns abstract grievances into specific, resolvable questions.

The throughline is the oldest one in this practice. The technology changes what review costs and how fast it runs; it does not change the neutral's obligation to hold the scales level. AI is a powerful instrument for reaching the relevant facts faster. It is also a powerful instrument for reaching them selectively. The neutral's task is to ensure the parties, and the record, can tell the difference.