FRESNO, Calif. – The Fresno Police Department is one of the first in California to use a new tool at the cutting edge of technology.
Unlike other commonly used police tech – including tasers, which deliver an electric charge to incapacitate a person, and body-worn cameras, which record encounters – this newest tool isn’t a device at all, but a form of artificial intelligence that comes to life on a computer screen.
It’s called Draft One, and it helps officers with the daily mundane task of writing police reports.
“The way that works is our officers will respond to their scene, they'll activate their body-worn camera, they'll record their video,” said Sgt. Steven Casto.
After the incident, he explained, Draft One transcribes the audio portion of that video and uses it to write a report. He refers to it as “an AI-powered narrative assistant.”
Fresno police started using Draft One in a limited number of incidents shortly before the company publicly debuted the product last April. In July, city leaders agreed to pay $1.3 million for the product as part of a new, 5-year police technology contract with Draft One creator Axon.
Fresno’s police department is the largest law enforcement agency in California using the tool.
Artificial intelligence in policing isn’t new. For years, AI has been changing how police do their work, from facial recognition screening to technology that tracks license plates.
But using AI to help write police reports is relatively untested. A limited number of other companies have recently rolled out similar technology, but none of them have Axon’s market share in police tech. The company is also the country’s sole manufacturer of tasers and the main provider of bodycams.
Axon touts Draft One as a “force multiplier for officers” – meaning it saves time and improves police reports without requiring more human resources.
At the same time, the police report is one of the most fundamental documents of the criminal justice system, and tasking AI with writing them raises concerns from tech experts, criminal defense attorneys and community members about accuracy and accountability.
How Draft One works
Draft One runs on the same technology that underlies ChatGPT, which uses conversational language to respond to prompts and produce essays on command. But where ChatGPT pulls its language from millions of sources of information, Draft One creates each police report based only on the audio from a single bodycam recording.
Casto and other members of the Fresno Police Department demonstrated the technology to KVPR in real time. Officer Gregory Colon-Reyes used a bodycam to record a portion of our interview, then generated a Draft One report using the footage. Within seconds, eight tidy paragraphs appeared on the screen.
“On September 24, 2024, at approximately 11:33 AM, I, Officer Gregory Colon-Reyes, was involved in a demonstration and discussion regarding the new body camera technology being implemented within our department,” the draft began.
The report was dry and factual, and included a striking number of precise details from the interview, including anecdotes from officers.
But this AI-produced document isn’t the end of the report-writing process: Officers must review each report, insert observations they made during the incident, and, finally, check a box that says they’re willing to testify to its accuracy.
Draft One does not generate its reports automatically after every incident. Officers have to choose to use the product.
Although the department has not formally analyzed time savings from its 2,000-plus Draft One reports generated so far, Casto said officers tell him they save as much as “20 or 30 minutes per report.” He said the writing quality is typically better, too.
So far, Fresno police have used Draft One in only misdemeanor cases. They plan to roll it out more widely, including for felony cases, as soon as early 2025. Interim Police Chief Mindy Casto – who is married to Sgt. Steven Casto – said she looks forward to the expansion.
“We're anticipating it being worth its weight in gold,” she said. “We're seeing very promising results so far.”
A study funded and conducted by Axon showed that reports written with the assistance of Draft One are of as good or better quality than those written only by officers. Axon’s website also boasts that the product “can save each officer an hour of paperwork each shift.”
Last month, however, the first peer-reviewed, independent study of Draft One suggested otherwise. Based on more than 700 reports produced over the course of a year by police in Manchester, N.H., the study, published in the Journal of Experimental Criminology, found no difference in report-writing time whether Draft One was used or not.
The study’s lead author, University of South Carolina criminology professor Ian Adams, said he and his colleagues are still looking into why the technology didn’t appear to save more time. He cautioned against relying too heavily on his team’s findings.
“It is in one agency, about one outcome at one point in time,” Adams told KVPR reporting partner KQED.
The ‘very worrisome’ potential problems
Some experts are concerned about Draft One’s accuracy, and are raising alarm about the potential for serious problems.
For Fresno defense attorney Kevin Little, that’s because of all the power held by police reports.
“A police report is used at so many stages of the criminal process,” he said, including in filing charges, determining whether a suspect is eligible for pre-trial release, and in writing bail reports and sentencing reports. “The fact that [police reports] may be computer-generated and not necessarily accurate is something that's very worrisome.”
Little also warns that a report based on bodycam audio may be incomplete. For instance, a camera records just one perspective in potentially chaotic situations — and it may not be recording at key moments. Plus, he asks: what if Draft One omits important information – or makes a mistake – and an officer doesn’t notice?
“Let's not pretend that this new technology is going to be any better, at the end of the day, than those who are controlling it,” he said.
Matthew Guariglia, a policy analyst with the non-profit Electronic Frontier Foundation, said the implications of a mistake – or even a misinterpretation – by Draft One are enormous.
“One of the big questions I have is: If an officer says into their body camera ‘drop the gun,’ will the ultimate report say ‘I said drop the gun,’ or will it say ‘the suspect was armed?’” Guariglia said in a recent interview with reporting partner KQED.
Those concerns were echoed by Sandra Celedon, President of the non-profit Fresno Building Healthy Communities and a member of the Fresno Commission for Police Reform formed in 2020.
“There's already a whole lot of reports out there that really talk about AI technologies really being racist, right, and really not understanding diversity and the dynamics of human interaction,” said Celedon, referring to a growing body of research suggesting the technology underlying AI can show biases based simply on a person’s name, gender or dialect. “It's a technology that is exploding right now, but we don't know enough about it to deploy it in a way that has the potential to imprison folks.”
The technology was built with safeguards to avoid those kinds of conflicts, said Axon product designer Noah Spitzer-Williams.
For instance, the personal observations that officers are required to make before signing off on the report are mandatory. Draft One inserts prompts that have to be completed before the report can be saved.
“We don't let the officers actually submit the report [without them],” he said. “It's a little bit of a speed bump to make sure that they're really doing the proofreading that we expect of them.”
Spitzer-Williams also said that Draft One isn’t nearly as imaginative as ChatGPT, which explains the just-the-facts dryness of the demo report that Colon-Reyes helped produce for KVPR.
“I call it the ‘creativity dial,’” said Axon’s Spitzer-Williams. “We literally turn it down to zero and it does make a dramatic difference. You see a noticeable drop in embellishments.”
Each report also includes a disclaimer at the very top that says it was written with the help of AI – even if the officer ends up rewriting most of the report.
“If you use a single word of that draft – or you simply use it as an outline, so to speak, but type all of your own narrative in its entirety – you’ve used [the technology],” he said. “If you use it in any manner, then it's a Draft One report.”
This story was produced with the support of the California Newsroom and with reporting contributed by KQED’s Sukey Lewis.
