How do you defend a predictive coding protocol when opposing counsel demands the seed set and training documents?

Predictive coding, a form of technology-assisted review (TAR), uses machine learning to prioritize and classify documents for relevance. When opposing counsel demands the seed set and training documents, the dispute is usually about transparency and defensibility, not the technology itself. A well-prepared producing party can respond from a position of strength.

Focus on Results, Not Internals

The core principle in most US civil matters is that discovery obligations attend to the adequacy of the production, not the mechanics of how you got there. A party generally must produce responsive, non-privileged documents and conduct a reasonable search. It is not ordinarily required to disclose the internal workings of its review process unless a deficiency is shown. Courts have increasingly recognized that demanding the seed set and training documents is a request to audit process rather than to test results.

Anticipate Privilege and Work-Product Concerns

Seed sets and training decisions often reflect counsel’s relevance judgments and mental impressions, which may be protected as work product. Producing them wholesale can waive protections or expose strategy. Note, too, that training documents frequently include non-responsive and privileged material that the requesting party would not otherwise be entitled to see.

Build Defensibility Before the Dispute

The strongest defense is established early:

• Validate statistically. Use sampling and recall/precision metrics to demonstrate the production captured responsive material.
• Document the workflow. Record the protocol, decision criteria, and quality-control steps contemporaneously.
• Cooperate proportionally. Meet and confer early; disclose your methodology (that TAR was used, the validation approach, and quality measures) even if you decline to hand over the seed set itself.
• Invoke proportionality. Discovery should be proportional to the needs of the case, and intrusive process discovery rarely meets that bar absent evidence of an inadequate production.

Practical Posture

Offer transparency about validation results and process, reserve privileged training inputs, and require the requesting party to articulate a specific, evidence-based deficiency before deeper disclosure. Rules and standards differ by jurisdiction (state courts and other countries), so confirm local rules and any governing TAR protocol.

For broader context, see e-discovery.