Reading Between The Lines: Modeling and Evaluating Behavioral Realism in Legal Simulation

A new legal simulation tool, WitnessSim, has been developed to enhance deposition training by simulating dynamic witness behavior. Unlike previous legal-AI evaluations focused on factual accuracy, this research introduces an evaluation framework that distinguishes between behavioral realism and pedagogical utility. The findings indicate that WitnessSim effectively maintains plausible behavioral boundaries and that legal professionals do not systematically distinguish between original and AI-generated testimony in terms of realism. Additionally, the tool demonstrated pedagogical usefulness, as witness behavior adjusted meaningfully based on question phrasing.

Why it matters

This development represents a significant advancement in the application of AI for professional training, particularly in fields requiring nuanced human interaction and dynamic response. It enables more realistic and adaptable simulation environments, potentially transforming how complex professional skills are taught and refined across various sectors beyond legal training.

Key insights

• WitnessSim is a deposition simulator designed to model dynamic witness behavior for attorney training.
• The evaluation framework used separates behavioral realism from pedagogical usefulness in legal AI.
• Realism was assessed through adversarial testing, blinded attorney comparisons, and analysis of behavioral trajectories.
• WitnessSim generally maintained plausible behavioral boundaries, suggesting a high degree of realism in its simulations.
• Attorneys did not consistently prefer original testimony over WitnessSim-generated testimony, indicating comparable realism.
• Pedagogical tests confirmed that simulated witness behavior adapted meaningfully in response to different question forms, demonstrating the tool's training utility.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.13712