Researchers reanalyzed a Redgrave LLP case study on eDiscovery tools. They found aiR and manual review perform similarly when method differences are considered. Continuous Active Learning models matched state-of-the-art effectiveness. These models also proved more efficient with limited training data. The prior comparison did not account for differing objectives and trade-offs. Inter-assessor disagreement further skewed the original results. The analysis used a sample of 2,000 reviewer determinations. The paper was submitted to arXiv on 12 October 2026.
What changed
A prior case study claimed aiR outperformed RAL and manual review without accounting for method differences.
What it unlocks
Comparing generative AI, machine learning, and manual review for eDiscovery on equal terms.
- 88% recall for aiR
- 64% recall for RAL
- 2,000 first-pass-reviewer responsiveness determinations
Sources