AI turns police bodycam footage into text. Can it keep the speakers straight?
AI police reports promise less paperwork, but getting the words right is only part of the job. Northeastern researchers explore why keeping track of who said what matters.
Turning people into frogs — and back into princes — is the stuff of fairy tales. But back in January 2026 this fate befell one Heber City, Utah policeman, when AI accidentally cast the spell. Transcription software picked up the dialogue from “The Princess and the Frog” playing at the scene where the police encounter was unfolding. When AI software used the recording to draft the report, it mistook the movie dialogue for part of the encounter and mistakenly claimed the officer had turned into a frog while on duty.
The mix-up is a reminder that AI might take busywork off your hands, but a human still has to be at the helm.
This is exactly what Northeastern professor of criminology and criminal justice Eric Piza and graduate student Savannah Reid found when they set out to assess how well AI transcription software captured dialogues recorded by body cameras. As they report in their recent paper, the bots whose work they reviewed got most of the text right, but struggled to keep the speakers straight when they switched turns.
The results show that human review remains essential before those transcripts are used to build a criminal case or hold an officer accountable, the researchers said. By pinpointing exactly where AI stumbles, the study helps investigators and researchers know what to check before relying on a transcript — for example, whether voices are missing or statements are attributed to the wrong speaker.
The software identified fewer participants per transcript in the AI-generated texts? and assigned longer passages to a single speaker, losing some of the back-and-forth. AI identified an average of 3.3 participants compared with 4.8 in the versions edited by people — and maxed out at seven, when its human counterparts identified as many as 16. AI transcripts were also watered down, containing about 25,000 fewer words across the sample. Missing dialogue could mean losing an explanation, warning or response that helps establish why an encounter escalated.
The researchers compared transcripts from 176 body camera videos that generated around 23 hours of footage covering 73 incidents. AI created one set, while human transcribers edited the other.
Ph.D.
Ph.D.
Ph.D.
The amount is a mere fraction of the more than 350 hours of footage generated by Kansas City police in a single day, the authors explained.
“Recording an encounter is much easier than finding the time to watch and analyze everything that gets recorded,” Reid said.
“We have turned supervisors into ‘movie watchers’ when they should be present and focused on developing the skills of their officers,” Shellie Solomon of Justice & Security Strategies, a research and consulting firm, added.
With little time to spare, the prospect of AI turning footage into searchable text is all the more appealing.
But getting the words down is only part of the job.
The bots also need to keep track of who said what — and sometimes, they got that wrong. There were no amphibian cameos among the samples, but one mix-up in particular raised some eyebrows. In a tense exchange between an officer and a civilian, the AI captured nearly all the words, but combined statements from different speakers.
“I ain’t selling no more dope,” the person stressed, insisting they weren’t going back to prison.
“You’re going,” the officer countered.
“I’m done selling PCP,” the person responded.
The AI transcript mistakenly made all three lines come out of the same mouth.
“Someone … could come away with the wrong impression,” Reid said.
Editor’s Picks
It’s easy to see where the confusion comes from, she added. Bodycam recordings aren’t made under studio conditions. Amid the chaos, even someone at the scene could struggle to tell speakers apart.
The Kansas Police Department did not immediately respond to a request for comment from Northeastern Global News regarding the findings of the study.
In the end, while camera footage — or the AI used to analyze it — won’t automatically make policing more transparent, it provides an independent log of events and draws attention to footage that might otherwise slip through the cracks.
A searchable record means public defenders can “find the two minutes they’re looking for without having to scrub through a three-hour video file,” Logan Seacrest, fellow at the R Street Institute, a public policy think tank, told NGN.
“The catch is that AI puts a layer of interpretation between the recording and the record,” he said, adding that due to a phenomenon called automation bias, people tend to give a machine the benefit of the doubt.
For that reason, it’s important to keep the roles clear.
“Think of (AI) as the investigative assistant who prepares the file,” Solomon explained. “The humans decide what the details mean.”
Reid also suggested checking what it leaves out of the review pile, since those encounters may never get a second look.
Missing words aren’t the only potential omission, however.
Key contextual details might not be part of the transcript at all, Calli Schroeder from Electronic Privacy Information Center, a public interest research center, noted. For example, say an officer accuses a peaceful citizen of acting aggressively. Will the system figure it out by pitting the text against the footage?
The camera itself captures only what’s in front of it, Solomon observed. It’s a device with a limited perspective rather than an “all-seeing presence.”
“AI that analyzes the footage has the same or perhaps more blind spots,” she added.
Seacrest described a Florida incident that puts this issue in clear view. A tiny acorn striking a patrol car sounded like a gunshot and sent a deputy diving to the ground in fear of his life as two officers fired into the police vehicle, where a handcuffed man miraculously escaped injury. A video taken from the right angle solved the mystery.
Ultimately, every small deviation could have profound implications, the researchers said.
“An AI transcription or summary may be 98% accurate, but that 2% could completely change who is at fault and what occurred in an encounter,” Schroeder explained.
Piza sees promise in future tools that could go beyond audio transcription to also analyze video patterns.
“I guess I’m a very cautious optimist about AI and policing,” he said.
