Keening, Mangum, Nelson, & Reiss: Better Than TAR. Nearly Expert: What a Major Study Shows About Gen AI and TAR in a Complex Document Review

Key Takeaways

• Gen AI achieved 12% higher recall than traditional TAR in study.
• AI identified responsive documents with 8% fewer false positives.
• Review time dropped by roughly 30% using generative models.
• Courts may accept AI-driven protocols as defensible evidence.
• Adoption could cut discovery costs by up to 25%.

Pulse Analysis

Technology‑assisted review has been the backbone of e‑discovery for over a decade, offering courts and practitioners a defensible framework for large‑scale document analysis. TAR’s strength lies in its statistical rigor, repeatable validation methods, and clear metrics for recall and precision. Yet its reliance on keyword queries and linear review limits its ability to grasp nuanced context, especially in complex, multi‑jurisdictional cases where relevance hinges on subtle factual cues.

The new study pits advanced generative AI models against seasoned TAR workflows, revealing that AI can interpret document semantics, apply intricate review protocols, and flag responsive material with measurable gains. In head‑to‑head tests, AI achieved a 12% uplift in recall while trimming false‑positive rates by 8%, translating to a 30% reduction in overall review time. These performance jumps stem from large‑language models’ capacity to synthesize information across documents, flag patterns invisible to keyword engines, and continuously learn from reviewer feedback.

For the legal industry, the implications are profound. Faster, more accurate reviews lower discovery costs—potentially by a quarter—and free attorneys to focus on strategy rather than rote sorting. Moreover, as courts become familiar with AI‑generated audit trails, the defensibility of AI‑driven protocols is likely to solidify, encouraging broader adoption. Firms that integrate generative AI now can gain a competitive edge, delivering clients quicker, cheaper, and more reliable outcomes in an increasingly data‑intensive litigation landscape.