The technical limitations highlight risks when AI-generated transcripts replace human reviews in criminal cases.
Eric Piza and Savannah Reid Evaluate AI Transcription Accuracy
Artificial intelligence tools designed to convert police body camera audio into searchable text capture a majority of dialogue correctly, but the software stumbles when multiple participants trade turns in a conversation. Eric Piza, a Northeastern University professor of criminology and criminal justice, and graduate student Savannah Reid assessed how well AI transcription captured dialogues recorded by law enforcement cameras.
Their research analyzed transcripts from 176 body camera videos, generating roughly 23 hours of footage across 73 separate incidents. Researchers compared text produced by artificial intelligence against versions edited by human transcribers to measure what the automated systems missed.
Speaker Tracking Failures and Dialogue Reductions
The evaluation revealed clear operational boundaries for automated transcription tools. AI software identified an average of 3.3 participants per transcript, whereas human editors recognized 4.8 participants. While human-edited versions identified as many as 16 speakers in a single session, automated bots capped out at seven.
The automated systems assigned much longer continuous passages to single speakers, erasing the natural back-and-forth rhythm of an encounter. Across the entire sample, AI-generated transcripts contained approximately 25,000 fewer words than the human-edited versions.
The software identified fewer participants per transcript and consolidated dialogue streams in ways that obscured who spoke specific warnings, explanations, or responses.
Supervisors Transformed Into Movie Watchers
The push toward automated transcription stems from sheer volume rather than administrative luxury. Kansas City police generate more than 350 hours of body camera footage in a single day, a scale that overwhelms traditional review methods. Recording an encounter is much easier than finding the time to watch and analyze everything that gets recorded,
Reid explained.
Outsourcing the initial review to software creates unintended organizational changes within police departments. Shellie Solomon, chief executive officer of Justice & Security Strategies, cautioned that automated workflows risk altering supervisory duties.
We have turned supervisors into ‘movie watchers’ when they should be present and focused on developing the skills of their officers.
Shellie Solomon, chief executive officer of Justice & Security Strategies
Transcript Gaps Threaten Criminal Case Integrity and Transparency
Missing dialogue carries profound legal consequences because dropped warnings or omitted responses can distort the official record of why an incident escalated. Logan Seacrest, a criminal justice and civil liberties fellow at the R Street Institute, noted that searchable records help public defenders find the two minutes they’re looking for without having to scrub through a three-hour video file.
Yet Seacrest pointed out the core vulnerability of automated recordkeeping. The catch is that AI puts a layer of interpretation between the recording and the record,
he said. Even a minor accuracy deficit can alter legal outcomes.