Arnold & Porter

While TAR is routinely used for responsiveness review, many attorneys still hesitate to use it for privilege review because it has not proven as effective at identifying privileged documents. Practitioners often default to linear, document-by-document privilege review by human reviewers, which is costly and time-intensive on large datasets.

The author used Relativity aiR for Privilege, a generative AI / LLM tool, to assist with privilege review in a large antitrust litigation document set of roughly 500,000 documents. Attorneys first spent nearly two weeks classifying law firms, attorneys, and third parties as aligned, adverse, or neutral so the LLM could assess privilege. The tool produced privilege predictions, category assignments, and rationales; humans then sampled and QC’d privileged, non-privileged, nuanced “in between,” and borderline documents.

Generative AI streamlined privilege review and provided useful categorization for targeting subsequent human review and QC, though it did not eliminate the need for human validation. Relative to the estimated cost of a full linear review with contract attorneys, the author achieved approximately 75% cost savings in this matter. Cost-effectiveness still depends on dataset size, Relativity’s per-document LLM fees, and how much human QC is required.

• • Generative AI privilege tools can deliver major cost savings versus linear review, but require substantial up-front attorney effort to map privilege-conferring and privilege-breaking relationships.
• • High-recall AI first passes still need human sampling and QC, especially for nuanced and borderline privilege categories.
• • Cost-effectiveness hinges on volume and QC intensity; per-document LLM pricing can erase savings on very large sets or when extensive human review remains necessary.