Imagine reading a transcript of a police encounter in which an officer and a civilian appear to say exactly the same things. The words are recognizable, the exchange looks orderly, and the document has the reassuring appearance of a factual record. There is just one problem: the computer has assigned both sides of the conversation to one person.
That is the kind of error Northeastern University researchers found while examining artificial-intelligence transcriptions of police body-camera footage. Their study raises an uncomfortable question at a moment when institutions increasingly use software to convert recordings into official-looking text: what good is an accurate quotation when it is attributed to the wrong person?
When a transcript changes the conversation
Criminology researchers Savannah Reid and Eric Piza compared AI-generated and human-edited transcripts from 176 body-camera videos, covering 73 police incidents and roughly 23 hours of footage. The automated system generally recovered much of the spoken language, but it had substantially more trouble determining who was speaking.
On average, the AI transcripts identified 3.3 speakers per video, compared with 4.8 in the human-edited versions. Across the sample, the automated transcripts also contained roughly 25,000 fewer words. In busy encounters, missing a bystander, merging voices or omitting an exchange can change how a reader understands what happened.
One example illustrates the stakes. During a discussion about drug dealing and prison, a civilian made statements about no longer selling drugs, while an officer interjected with a warning that the person was going back to prison. The AI transcript combined the remarks into a single speaker’s account. It preserved much of the vocabulary while distorting the relationship between the people talking.
A camera records; a transcript interprets
Body-worn cameras are intended to preserve evidence of encounters that might otherwise depend on competing recollections. But a video is not the same thing as a transcript, and a transcript is not the same thing as a conclusion. Someone—or something—must decide where one speaker stops, another begins and which sounds deserve inclusion.
Those decisions are difficult even for people. Police recordings capture overlapping voices, movement, background noise and moments of confusion. Automation could make hours of footage searchable and help investigators, defense attorneys and supervisors locate important passages. The researchers do not argue that those uses should disappear. Their findings instead draw a boundary between a tool for finding material and a record trusted to establish responsibility.
The study examined Rev.ai transcription output from one large U.S. police department. Its results should not be treated as a measurement of every AI product or every police agency. But the failure mode is consequential because the format looks authoritative: a clean transcript can hide the uncertainty of the recording behind it.
The New Jersey question
The implications reach New Jersey, where body-worn camera recordings are part of the law-enforcement accountability framework. State law addresses officers’ review of footage and requires disclosure in specified circumstances when an officer has reviewed or received an account of a recording before preparing a report or statement.
What has not been established by this Northeastern study is whether New Jersey state, county or municipal agencies are currently using comparable AI-generated transcripts or summaries, and if so, for which purposes and under what review rules. Those are questions worth asking departments and prosecutors directly: Are automated transcripts used to search footage, prepare reports or inform evidence review? Who checks speaker attribution? Is the original video available whenever the text is cited?
Accuracy is not the same as truth
Across workplaces, newsrooms and public institutions, AI-generated notes and summaries are becoming ordinary. Their efficiency is appealing because they transform an unwieldy recording into something readable in seconds. The risk is that readers can begin treating the polished output as an independent account rather than a machine’s interpretation of source material.
In a meeting, misattributing a comment might create an awkward disagreement. In a police encounter, it could affect how a warning, admission or accusation is understood. The answer is not necessarily to abandon transcription technology, but to keep the underlying footage accessible and ensure meaningful human review when the stakes are high.
The body camera was supposed to help answer what happened. If its transcript changes who said what, the record can create a new dispute of its own.
Sources & reporting notes
Northeastern University, “AI turns police bodycam footage into text. Can it keep the speakers straight?”, October 8, 2026. https://news.northeastern.edu/2026/10/08/ai-body-camera-transcription/
Savannah Reid and Eric Piza, “AI vs. human transcription: evaluating accuracy and meaning in police body-worn camera footage,” Journal of Experimental Criminology (2026). https://link.springer.com/article/10.1007/s11292-026-09774-0
New Jersey Public Law 2023, Chapter 177 (body-worn camera provisions). https://pub.njleg.state.nj.us/Bills/2022/PL23/177_.HTM
Reporting note: No claim is made that any specific New Jersey agency uses AI transcription; this would require agency-level confirmation. Human-edited transcripts, rather than independently produced verbatim transcripts, were the study comparison.
Northeastern University, “AI turns police bodycam footage into text. Can it keep the speakers straight?”, October 8, 2026. https://news.northeastern.edu/2026/10/08/ai-body-camera-transcription/