Researchers at Vanderbilt Law School have proposed new policies to govern the use of artificial intelligence for reviewing police body-camera footage. The push comes as departments increasingly turn to automated tools to manage the massive volume of video generated by officers.
Automating the review process
Following the 2014 shooting of Michael Brown, the U.S. Department of Justice supported the adoption of body-worn cameras. These devices record interactions and are intended to create a psychological loop that deters aggression. By 2020, roughly 79 percent of U.S. police officers worked in departments that used the technology. The widespread adoption has produced more than 5,000 years’ worth of video. Manually reviewing this footage is practically impossible for human analysts. Recently, however, several AI vendors have developed tools to automate the review process.
The Vanderbilt team evaluated one such tool, Truleo, which extracts audio from police footage to create a transcript. The system separates the interaction by speaker and classifies it by type, such as traffic stops or stop-and-frisk encounters. It also parses police speech for insults, threats, profanity, and the threat of force. When an officer uses restraint, Truleo applies positive tags. Conversely, the system flags interactions involving insults or threats as “pending.” These flagged items are sent to police supervisors, who may apply a “follow-up” label or remove the flag from the review workstream.
The researchers note that Truleo’s audio-to-text system minimizes the risk of abuse by removing identifying visual markers. This approach prevents linking body-worn camera data to specific police investigations. The system also includes automatic personal information redaction and search limitations to limit surveillance risks. Despite these design choices, the scholars express concerns about the influence that police departments have over AI companies. Police associations are primarily concerned with the welfare of their officers and have historically resisted accountability efforts. Market pressure already shapes Truleo’s policies. The team documented a shift in the company’s marketing from a focus on police accountability to one promoting “police professionalism” and morale.
Related: Complaining Neighbor Gets Half a Garage After Dispute
Heydari and his coauthors observed that Truleo’s design requires human intervention to apply negative labels, which limits accountability. Pressure from police unions reportedly led the company to reconfigure its system to suppress automatic supervisor notifications about officers with higher-than-normal negative interaction rates.
Risks of automated labeling
While the technology offers a way to process vast amounts of data, it carries significant risks. The Heydari team warns that AI labels may miss the mark by prioritizing superficial criteria that do not necessarily imply good policing. For example, a pretextual stop may remain superficially professional and courteous, even if motivated by unfair factors. The system would likely not flag such interactions. Positive labels may also anchor reviewers to a positive framing of police interactions, warping data and analysis. Police departments may be tempted to mischaracterize data by touting high rates of positive tags as evidence of effective policing.
Even with these limitations, the researchers argue that AI-enhanced camera data review can support meaningful law enforcement reforms if implemented correctly. To prevent misuse, they offer concrete guidelines for police departments and AI developers. The scholars advocate adopting evidence-based labels and suggest that vendors regularly test these categories against independent human control groups. They also urge both agencies and vendors to provide the public with clear definitions of data classifications so that stakeholders understand what behaviors the system measures. This approach helps ensure that the tools function as intended rather than masking underlying issues.
Ensuring transparency is a central part of the proposed framework. Vendors should provide independent verification of any claimed benefits, and companies must clearly define their metrics. Police departments should be transparent about how they implement the software. At a higher level, the team advocates for the municipal ownership of body-worn camera data or robust data-sharing policies. This approach would ensure that prosecutors and defense attorneys have ready access to evidence and unlock the reform value of footage archives by enabling independent researchers to analyze it at scale.
