Brief

Human-on-the-Loop Orchestration for AI-Assisted Legal Discovery

A new paper proposes a four-layer verification architecture to mitigate errors in AI-assisted legal discovery. The proposed system aims to prevent "trajectory collapse," where early misclassifications in autonomous LLM agents lead to legal malpractice. A simulation study on a synthetic e-discovery corpus demonstrated that mandatory Human-on-the-Loop escalation thresholds can reduce privilege-waiver risk by up to 61% compared to fully autonomous systems, while still routing less than a quarter of documents for attorney review. AI

IMPACT Introduces a framework to reduce legal risks associated with autonomous AI agents in sensitive document review processes.

• Hugging Face
• arXiv
• large language model
• DagsHub
• alphaXiv
• CORE Recommender
• ScienceCast
• CatalyzeX
• Gotit.pub
• electronic discovery
• Influence Flower
• Human-on-the-Loop