{"slug": "millions-of-hours-of-bay-area-police-body-camera-footage-go-unwatched-stanford", "title": "Millions of hours of Bay Area police body-camera footage go unwatched. Stanford says AI could change that", "summary": "Stanford University researchers have developed an AI tool to analyze millions of hours of police body-camera footage, starting with a study of nearly 1,800 NYPD stops that revealed officers often failed to properly seek consent for searches and showed racial disparities. The tool is now being applied to over 300,000 hours of footage from two Bay Area departments, including Oakland, promising unprecedented oversight of police interactions.", "body_md": "**Getting your**\n\n[Trinity Audio](//trinityaudio.ai)player ready...The promise for more than a decade as police departments adopted body cameras was a new era of accountability, where footage would provide a clearer record of an interaction between officers and the public.\n\nBut in practice, most body-camera footage is never closely reviewed. The public usually sees it only after the most serious or high-profile encounters, just a sliver of everyday police interactions.\n\nResearchers with Stanford University’s [SPARQ behavioral science team said they have made a breakthrough](https://sites.google.com/stanford.edu/ai-policing-study) in answering that question, with analysis of police interactions in New York leading to a new program looking at Bay Area departments. They used artificial intelligence tools to comb through a sample of nearly 1,800 New York police street stops recorded on officers’ body cameras to assess whether the officers were properly seeking consent before searching people, as part of a 2021 court-ordered study to measure the impact of reforms to the police department’s “stop and frisk” practices.\n\nThe findings point to possible noncompliance, underreporting and racial disparities. But they may end up being secondary to the broader potential the study represents. Millions of hours of officer body-camera footage — audited via relatively tiny samples until now — could be meaningfully analyzed, providing entirely new levels of insight into how officers are abiding by constitutional obligations in their everyday work.\n\n“I feel like for the first time, we can use AI to analyze millions of police interactions at scale,” said Jennifer Eberhardt, a Stanford psychology professor and faculty co-director of the SPARQ team. “This knowledge offers an unprecedented opportunity to understand and also to improve policing not just in one department, but across the country … It’s a game changer in that way.”\n\nResearchers said the model — detailed in a report published this spring — is being applied in two large cities in the Bay Area, involving a data set of more than 1.3 million body-camera videos covering over 300,000 hours.\n\nOne of them is the Oakland Police Department, which participated in previous Stanford research where footage was analyzed to assess officers’ level of respectfulness during traffic stops and how that intersected with race as well as attempting to determine the chances of the stop escalating into serious encounter or violence. The second agency receiving the new form of AI analysis was not disclosed, with the university citing that their work is unfinished and that the organization does not have a public relationship with the research team like in Oakland.\n\nThe new study shows how AI could help deliver the oversight many expected when police body cameras became widespread, said Rob Voigt, a UC Davis assistant professor of linguistics and SPARQ affiliate.\n\n“The overwhelming majority of body camera footage that is recorded around the country, millions and millions of recordings of police interacting with people around the country every year … nothing happens to it. It’s only a piece of evidence when something goes wrong,” Voigt said. “We’ve been working for more than a decade under this idea that the footage is data, and that we can use it to understand these interactions at scale.”\n\nThe NYPD analysis involved taking transcripts generated from footage recorded on Axon body cameras and running them through machine learning and natural language processing tools. The results indicate that officers were often sidestepping explicitly asking for search consent, in favor of veiled and passive language that was difficult for people to understand in real time. The research team also found troubling racial disparities when officers used language reserved for serious suspicions — like phrasing implying commands or accusations — when the reason for the stop was relatively minor.\n\nIn the sample of recorded encounters from 2023, the study found that when NYPD officers asked to search someone, they used the word “consent” in 12.7% of searches that were documented as consent searches, and the word “search” 46% of the time.Instead of direct language, officers asked if they could “check” someone in 36.7% of instances, and used the phrase “Do you mind?” in 16.8%. The study argues that such indirect language obscures a subject’s right to refuse a search and could violate their Fourth Amendment rights.\n\n“Even though these people, from a legal perspective, are free to go, there’s always a question of how much does the person understand that they are free to go,” Voigt said. “To a significant degree, they are hearing language that is more like stops.”\n\nThe study also found racial disparities in instances when officers used detention-oriented language in these lower-level “free to go” encounters, with Hispanic and Black people having a “stop probability” of 5% to 11% higher than white people even after controlling for local crime rate. With higher-level stops, the disparity held for Black people, at between 5% and 17% higher than white and other races.\n\nThe AI analysis model also examined an estimated 25% underreporting rate of street stops described in a previous academic study. Some of the undocumented stops were found to be misreported as “low-level encounters” where no detention occurred. Eberhardt said this illustrates a discord in which officers are failing to document stops that they know are being recorded, adding that officers have to know the chances of their footage being examined is “pretty low.”“In one year alone, the NYPD recorded 4.7 million encounters but the monitor reviewed fewer than 2,500 of those,” she said. “Having those cameras doesn’t have an impact on police officer behavior in the way that you might want.”\n\nThe disparity between documented and actual consent search stops is significant, the study contends, because noncompliance is more present in the undocumented instances. Officers are more likely to state a reason for a documented stop and make clear whether the subjects are free to go, while the stops that were recorded on camera but never properly classified more frequently include “civilian pushback” and more indirect and veiled consent language.\n\n“There are a lot of questions and dark matter in body-worn camera analysis. What don’t we know?” Voigt said. “It puts a much finer point on underreporting as this huge issue. How can we fix that?”\n\nAt several points the study acknowledges its limitations, including the analysis being restricted to vendor-generated transcripts of officer footage and the relatively small sample size. Dylan Verner-Crist, lead senior investigator with the ACLU of Northern California, said the goal should ultimately be having access to untouched data.\n\n“If you use an automated transcriber, it’s decent for a one-on-one conversation,” he said. “But for a body cam recording on the side of the road, where there’s often loud traffic noise, wind, things like that. Automated transcription can be sort of rife with problems.”\n\nHe also emphasized that police agencies need to make massive strides in body camera footage accessibility to scale it out, noting that the Oakland research work and the recent NYPD study were built on data samples obtained through complex negotiations or a court order.\n\n“It might technically be public record, but that doesn’t mean you can get it, or doesn’t mean you can get it in any reasonable time frame,” Verner-Crist said.\n\nThe AI analysis is proof of concept of a workable way of digesting massive stores of data to get answers, according to the authors, which along with Eberhardt and Voigt include Dan Sutton, director of justice and safety at the Stanford Center for Racial Justice at Stanford Law School, and University of Michigan assistant professor Nicholas Camp, a longtime SPARQ collaborator.\n\n“It’s a lightweight way to use the information that’s already there to ask really important questions and identify big issues,” Voigt said.\n\nVerner-Crist agrees with the promise shown by the work, saying it reflects a “doable chunk” of AI analysis where the technology “helps me identify the stops that I need to watch,” but still requires thorough human review and evaluation. It’s an example, he says, of how AI and its precursors are additive at best, and that no one should expect the tech to displace human judgment and insight.\n\n“You need to know case law. You need to know the Fourth Amendment to make the analysis, and I can’t see a world where we have AI making analyses like that,” he said.\n\nThe authors also frame the revelations from their findings as being geared toward corrective measures and early intervention, “to select potentially problematic footage for human review and substantially raise identification rates of undocumented stops.”\n\n“We have an opportunity to really look closely and to not only see whether those interactions improved generally, but to know what aspects of the training might actually be really landing with officers,” Eberhardt said. “We have this tool now where somebody’s watching what’s going on … the presence of footage alone is not doing the intervention work that we thought.”", "url": "https://wpnews.pro/news/millions-of-hours-of-bay-area-police-body-camera-footage-go-unwatched-stanford", "canonical_source": "https://www.mercurynews.com/2026/06/17/millions-of-hours-of-bay-area-police-body-camera-footage-go-unwatched-stanford-says-ai-could-change-that/", "published_at": "2026-06-17 11:30:39+00:00", "updated_at": "2026-06-17 11:55:21.393148+00:00", "lang": "en", "topics": ["artificial-intelligence", "natural-language-processing", "ai-ethics", "ai-research", "machine-learning"], "entities": ["Stanford University", "SPARQ", "NYPD", "Oakland Police Department", "Axon", "Jennifer Eberhardt", "Rob Voigt", "UC Davis"], "alternates": {"html": "https://wpnews.pro/news/millions-of-hours-of-bay-area-police-body-camera-footage-go-unwatched-stanford", "markdown": "https://wpnews.pro/news/millions-of-hours-of-bay-area-police-body-camera-footage-go-unwatched-stanford.md", "text": "https://wpnews.pro/news/millions-of-hours-of-bay-area-police-body-camera-footage-go-unwatched-stanford.txt", "jsonld": "https://wpnews.pro/news/millions-of-hours-of-bay-area-police-body-camera-footage-go-unwatched-stanford.jsonld"}}