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

Summary

A study introduces WitnessSim, a deposition simulator built around controllable legal personas, designed to model and evaluate behavioral realism in legal simulation. The authors note that deposition training requires attorneys to handle dynamic witness behavior, yet existing legal-AI evaluations mostly focus on factual accuracy, reasoning or response-level plausibility rather than behavioral realism. They introduce an evaluation framework that treats behavioral realism and pedagogical usefulness as separate concerns. The paper assesses realism through adversarial testing of the simulator. The framework could become a new standard for verifying the reliability of AI tools used in legal training.