Whether you’re in-house at a corporation or running discovery at a firm, the question about Generative AI (GAI) isn’t whether it belongs in your workflow anymore. It’s how to use it without having to rebuild everything you already know about defensibility.
A new piece in Law.com’s Legaltech News, Ground Truth: The Realities of Generative AI in E-Discovery, takes that question on directly. eDiscovery AI Founder and CEO Jim Sullivan co-authored it with Esther Birnbaum (HaystackID), Michallynn Demiter and Ben Sexton (JND eDiscovery), and Cristin Traylor (Relativity), with contributing perspectives from practitioners at Microsoft, Walgreens, Morgan Lewis, Fidelity National Financial, Redgrave LLP, Purpose Legal, and Aligned Discovery.
The takeaway: GAI in eDiscovery doesn’t need a new rulebook. It needs practitioners who understand the tools and apply the rules already in the books.
Classification vs. Interpretation: Two Different Conversations
Not every GAI task is the same. Classification work (i.e., relevance, privilege, issue coding) is binary and measurable. Interpretive work (i.e., summaries, timelines, fact extraction) is open-ended.
Lumping them together is what’s muddied the validation conversation. Pulling them apart is how it gets simpler.
Classification Fits Inside Existing TAR Frameworks
For classification, the validation playbook already exists: recall, precision, elusion, and sampling. The same metrics courts have accepted since Da Silva Moore in 2012. The algorithm changed; the discipline didn’t. Across the case studies the article gathers, validated recall regularly lands above 90%, often above 95%. Quinn Emanuel reported a median estimated recall of 99.4%. Morgan Lewis used GAI on SEC and DOJ subpoena responses with the agencies’ advance approval.
For interpretive outputs, the rulebook is a professional responsibility. Attorneys are accountable for what they sign their name to. Automation bias is a real risk, but the answer lies in training and judgment, not in a new governance layer.
The proof-of-concept phase is over. The question is no longer whether GAI works for review. It’s how teams put it to work.
Read the full article on Law.com
Want to see what a defensible GAI review looks like? Schedule a demo
