Off-campus UNL users: To download campus access dissertations, please use the following link to log into our proxy server with your NU ID and password. When you are done browsing please remember to return to this page and log out.
Non-UNL users: Please talk to your librarian about requesting this dissertation through interlibrary loan.
Measuring Jury Perception of Explainable Machine Learning and Demonstrative Evidence
Abstract
Subjective pattern comparison has been subject to increased scrutiny by the courts and by the general public, resulting in an increased interest in pattern comparison algorithms that provide quantitative assessments of similarity for use by forensic scientists. While these algorithms would mark an improvement over current subjective comparison methods, individuals without a statistical background may struggle with the statistical concepts and language necessary for describing algorithmic methods. If algorithms are to be used, examiners must be able to testify about their use in a way that is accessible to the jury. In a series of studies, we conduct an assessment of language and supporting visual aids which might be used to explain bullet matching algorithms. In the initial study, we encountered a response type calibration issue - individuals thought highly of the forensic witness and evidence regardless of experimental conditions, ‘maxing out’ Likert response scales and leaving us unable to tell if the conditions had any effect. While this study indicated that individuals overall found the testimony to be reliable, credible, and scientific, it did not readily provide information about our question of interest. Additional data from this study was found in the participants’ note pads. Through cleaning sequential notes and designing a method for highlighting study transcripts according to the frequency of collocations in participant notes, we can determine which portions of testimony participants found ‘noteworthy’. We also conducted a study on response types to determine the consistency of participant responses across response types, compare a variety of response types, and determine which response type may be appropriately calibrated for addressing the initial research question of jury perception of algorithms and demonstrative evidence. The response types used in this investigation include the participant’s interpretation of the strength of evidence (Likert scale), conviction decision (binary), opinion of guilt (binary), willingness to bet on their opinion of guilt (numeric), probability of guilt (numeric), and chance of guilt/innocence (numeric or multiple choice). The note cleaning, text analysis, and testimony tools we developed throughout this series of experiments will benefit our future research in jury perception, as well as future transcript studies.
Subject Area
Statistics|Computer science|Artificial intelligence|Language
Recommended Citation
Rogers, Rachel Edie Sparks, "Measuring Jury Perception of Explainable Machine Learning and Demonstrative Evidence" (2024). ETD collection for University of Nebraska-Lincoln. AAI31240449.
https://digitalcommons.unl.edu/dissertations/AAI31240449