Draft One: Leveraging LLM to Convert Body-Cam Audio into Police Reports

GPT-4 Summary:

Axon has introduced Draft One, a pioneering AI-powered software that significantly enhances the efficiency of police report writing by using auto-transcribed audio from body-worn cameras. This innovation not only saves officers approximately one hour per day by reducing the time spent on report writing but also improves the quality and clarity of the reports. Draft One incorporates rigorous safeguards and requires human review to ensure accuracy and accountability, aligning with Axon’s commitment to responsible AI usage.

Draft One is designed to assist law enforcement agencies by automating the drafting of report narratives, thereby allowing officers to focus more on community engagement and other critical tasks. The tool has shown promising results in agency trials, significantly reducing the time officers spend on paperwork. Furthermore, the software adheres strictly to the audio data from recordings, ensuring that reports are factual and free of embellishments.

Axon’s approach emphasizes ethical AI development, incorporating feedback from its Ethics and Equity Advisory Council and conducting thorough bias testing to ensure the technology serves as a force for good. This initiative is part of Axon’s broader mission to enhance public safety and reduce gun-related deaths through innovative technology. Draft One is now available within the Axon ecosystem and is compatible with Axon body-worn cameras and Axon Evidence customers.

Draft One includes a range of strict safeguards and requires that humans make the decisions in key moments:

  • Officer required to review and sign-off: Once the report narrative has been edited and key information has been added, officers are required to sign-off on the report’s accuracy before submitting for the next round of human review.
  • Adheres to the audio data: Report narratives are drafted strictly from the audio transcript from the body-worn camera recording. Axon calibrated the underlying model for Draft One to prevent speculation or embellishments.
  • Controls to ensure proofreading: Each draft includes placeholders that officers are required to edit, by either adding more information or removing the placeholder.
  • Report drafts are restricted to minor incidents and charges: The default experience limits report drafts to minor incident types and charge levels, specifically excluding incidents where arrests took place and for felonies. With this default, agencies can get started with Draft One and gain tremendous value in expediting report writing for the overwhelming majority of reports officers submit. This allows agencies the option to gain experience on low severity reports first, then expand to more severe reports once they gain experience in how to use the tool effectively. Agencies can set a policy determining which reports are eligible for Draft One utilization, and the tool then ensures enforcement of the agency policy.

I wonder if humans will continue checking the reports for hallucinations… We all know how lazy some people can be when copying/pasting from LLMs…

1 Like