AI can now play a convincing deposition witness in blind tests with real lawyers.
The tool, called WitnessSim, generates witness testimony driven by controllable "legal personas" - profiles meant to capture how a certain type of witness behaves under questioning. The researchers evaluated it three ways: adversarial probing to see whether the persona would collapse under pressure, blinded comparisons where attorneys judged real testimony against simulated testimony, and tracking of how witness behavior shifted over a long line of questioning. WitnessSim generally held its assigned persona, and attorneys showed no consistent preference for the original testimony over the simulated version. The simulated witnesses also changed their answers meaningfully in response to different question phrasing and attorney tactics, rather than folding into one generic, cooperative AI voice.
Most legal-AI research so far asks whether a model gets facts or law right, not whether it behaves like a person under cross-examination. Deposition prep is expensive partly because it needs a live human sparring partner, usually another attorney or a paralegal standing in as witness. A tool that holds a consistent, pressure-tested persona could make that kind of practice cheaper and more available to smaller firms, which changes who gets to rehearse this skill at all, not just who gets replaced.
The paper does not claim WitnessSim is ready to sit in for a real mock witness. It claims something narrower: that lawyers, in blinded tests, could not reliably tell the difference, a bar most chatbot benchmarks never bother clearing.