Generative AI in eDiscovery: Defensibility, Risks, and Best Practices

Large language models are now embedded in mainstream review platforms, used for issue coding, privilege triage, deposition preparation, and document summarization. They promise dramatic efficiency gains. But a court does not care how a result was produced unless the proponent can show the result is reliable — and generative models introduce failure modes that traditional TAR did not. The defensibility question is whether you can audit and validate the output, not whether the technology is impressive.

The three risks that matter in court

Hallucination and fabricated support

A generative model can produce fluent, confident output that is simply wrong — a summary that misstates a document, or a citation to material that does not exist. In review, an unverified AI privilege call or issue tag is a defect waiting to be exposed. Every consequential output needs a human-in-the-loop verification path and a sampling regime that measures the model's actual accuracy on the corpus.

Confidentiality and data governance

Feeding client documents to a model raises immediate confidentiality, privilege-waiver, and data-residency questions. Litigators should confirm that processing occurs in a closed, contractually protected environment, that inputs are not used to train external models, and that the arrangement is consistent with protective orders and any cross-border data restrictions.

Opacity and auditability

A model that classifies in ways no human can fully explain raises the validation bar rather than lowering it. If you cannot trace why a document was tagged, you must prove adequacy statistically — through rigorous sampling of both the responsive and the discarded sets — to a degree that satisfies a skeptical adversary.

Best practices for a defensible workflow

1. Document the workflow before you run it — the model used, its role, the prompts or instructions, and the human review layered on top.
2. Validate with measurement — apply the same recall, precision, and elusion discipline used for TAR, with stated confidence levels and error margins.
3. Keep humans accountable for consequential calls — privilege and responsiveness determinations should be confirmed, not delegated wholesale to the model.
4. Lock down the environment — closed processing, no external training on your data, and contractual protection consistent with the protective order.
5. Negotiate AI use in the ESI protocol — disclose the role of generative tools and the validation approach rather than litigating it after production.

What courts and standards bodies expect

Guidance from The Sedona Conference and the broader bench in 2026 is consistent on a single theme: generative AI is permissible where its use is reasonable, documented, and validated. The technology does not change the governing standard — a reasonable, well-documented process — it raises the stakes for proving you met it. Courts have grown impatient with both reflexive resistance to AI and uncritical reliance on it.

The role of the expert

When generative AI is part of a production, an independent expert who can explain the workflow in plain terms, validate the output with defensible sampling, and withstand cross-examination on the model's limits is what separates an efficiency gain from an exclusion risk. The same Rule 702 discipline that governs any expert opinion governs here.

If generative AI is shaping your review or you are facing a challenge to an opponent's AI-assisted production, scope the engagement early. Start a conversation through our home page or email the team to discuss a conflict check and approach.