AI vs. human transcription: evaluating accuracy and meaning in police body-worn camera footage

The current study measures whether AI-generated transcripts of body-worn camera footage accurately capture who is speaking and what is said, compared to human-edited transcripts. We analyzed AI and human-edited transcripts from 176 body-worn camera videos (73 incidents, approximately 23 h of footage) from a large U.S. police department. We used multiple text-similarity metrics to compare speaker extraction, speaker matching, and overall document similarity. AI-generated transcripts captured much of the same lexical content as human-edited transcripts but performed noticeably worse on identifying and attributing speakers, especially in complex, multi-speaker encounters. AI transcription appears suitable for large-scale content search and descriptive text analysis but should not replace human transcription for attribution-sensitive uses. The findings underscore the value of hybrid human-AI workflows and the need for clear policies distinguishing low-stakes from high-stakes uses of AI-generated transcripts in policing.