Intelligent systems for uncovering truth and inconsistency in courtroom testimony

Cross-examinations and court hearings are a standard method for surfacing additional context and aiding in the interpretation of existing evidence during litigation. Lawyers currently follow a very labor-intensive process in validating claims made during witness hearings by manually screening the case record for any conflicting evidence. Locating the right evidence and subsequently drawing conclusions presents significant challenges, particularly within large-scale case-related document corpora that can easily grow to ten thousand pages in complex litigation and arbitration cases. We are developing an automated LLM-powered pipeline to locate conflicting and supporting evidence within legal documents, as well as detect deceptive behavior, in hearings. Deceptive behavior includes commission (active deception), omission (passive deception), and paltering. We are also building a prototype of a system that provides legal professionals with a reasoning trace and the supporting and conflicting evidence, assisting them in determining the correctness of statements made by the witness.

Additionally, we plan to release a public dataset of labeled hearing excerpts to the research community, thereby facilitating further advances in our newly defined task of detecting deceptive behavior in witness hearings.