○Police officer Wendy Venegas gently spoke to a 14-year-old boy in Spanish as he stood on the side of the road in a small residential area in East Palo Alto. The girl’s face was swollen from crying so much, and she calmly explained what had happened.
The Guardian’s journalism is independent. If you buy something through an affiliate link, we may earn a commission. learn more.
Venegas later explained that the girl’s father had caught her and her boyfriend “doing something” that morning, and that the father had either hit or shoved the boy. Well, the police arrived and questioned all three people. Everything up to this point was standard procedure.
But when it came time to compile the case into a report, Venegas had the help of a new assistant: a cutting-edge artificial intelligence tool called Draft One.
East Palo Alto is a small, working-class city that feels worlds apart from nearby Silicon Valley, and several California counties have begun using or testing AI, including Campbell, San Mateo, Bishop, and Fresno. . It is powered by software developed by Axon, an industry leader in body cameras and Tasers. Axon said the program will allow police officers to produce more objective reports in less time. But as more agencies deploy these types of tools, some experts are wondering whether we are giving artificial intelligence too much of a role in the criminal justice system.
“We forget that those documents play a truly central role in decisions that change people’s lives,” says American University in Washington, who wrote the first law review article on AI-assisted police reports. said Andrew Ferguson, a professor of criminal law at the College of Law. It is scheduled to be published next year.
From recording the details of complex murders to basic records of stolen bicycles, police reports are central to police work.
“These are actually kind of the building blocks of the criminal justice system because they’re the official commemoration of what happened, when it happened, and sometimes why,” Ferguson said.
Prosecutors make charging decisions, judges make bail decisions, and people make decisions about their own defense based, at least in part, on what is written on that first piece of paper.
“If some of that is being shaped by AI, that raises big concerns about whether we can rely on it,” Ferguson said. The potential for error and bias introduced by AI is still being studied. But he added that law enforcement leaders have a natural desire to improve efficiency.
Axon is pitching its “Draft One” tool as a force multiplier, appealing to many police departments struggling to recruit and retain officers, many of whom are law enforcement agencies in the wake of George Floyd’s killing. The crisis was exacerbated by the protests that followed.
“The pendulum has swung, but many companies are still 15 to 20 percent below their target strength numbers,” Axon founder Rick Smith said during an August earnings call with shareholders. “So you hear this kind of magical feedback: ‘If Draft 1 frees up 20 to 25 percent of an executive’s day from reporting, that’s almost like 20 days. ‘My Force power increased by % overnight.’
East Palo Alto Director Jeff Liu said his agency is not immune to these staffing concerns. The department has budgeted for 36 sworn officers, including the secretary, but he is currently short of eight positions. He doesn’t see Draft One as an immediate solution, but he hopes it will help officers spend more time on the street.
“If this AI speeds up reporting without compromising accuracy, I think that’s a win,” he said. According to East Palo Alto’s contract, Draft One will cost about $70 per month per body camera, or about $40,000 annually.
Liu said that while he hasn’t been writing much reporting lately, he has been using ChatGPT to draft social media posts and condolence letters and customize them with his own voice. He said using the popular AI chatbot made him more open to draft one. Draft One uses the same underlying AI as ChatGPT, but the department’s data is stored in a secure government cloud service developed by Microsoft.
Axon isn’t the only company offering this service. Truleo – a company that uses AI to analyze vast amounts of body camera footage to ensure officers are acting professionally – offers a similar reporting program, but not as good as Axon’s Draft One. It is not widely marketed or adopted.
In nearby Santa Clara County, Campbell Police Lt. Ian White said that during the first month his department tested Draft 1, officers said they saved about 50 hours overall. The Fort Collins, Colorado, Police Department found that reports created with Draft One took an average of 8 minutes to complete, compared to 23 minutes for reports not using the software.
However, the first independent study of the Draft One software, published this week in the Journal of Experimental Criminology, did not support the time savings reported by White et al. Researchers at the University of South Carolina conducted a randomized controlled trial with the New Hampshire Police Department over the past year and found that officers who used Draft 1 wrote reports faster than officers in the control group. It turns out there wasn’t.
Assistant Professor Ian Adams, who led the study, said he could not yet draw any conclusions about why the time savings were not achieved and said he and his colleagues were still investigating, but he was surprised by the results. He also cautioned against placing too much weight on his findings. “It’s done in one institution and for one outcome at a time,” Adams said.
His team is still studying whether there are other benefits, such as accuracy or completeness. Technology like Draft One could produce better-written reports that “could be returned for edits and revisions less frequently, resulting in overall system savings.” “You might see it,” he said.
Officer Venegas of East Palo Alto said the program helps him overcome writer’s block, especially after a long day on patrol. She simply presses the Draft One button on her computer, and within seconds a narrative based on an audio transcription of the body camera footage appears.
“You don’t know what words you’re going to write, you just look at them and you think, ‘Oh, that’s exactly what I was thinking!'” That’s the best part, she says. said. Draft One is also changing the way people work in the field. Because the report is based on audio recordings, Venegas said, it intentionally speaks about what is happening during the incident.
“I would say, ‘Did you see that? “The mirror is broken,” Venegas said. “‘Did you see? There’s something on the floor. A knife, a bloody knife on the floor.”
Axon product designer Noah Spitzer Williams said this is one of the software’s most surprising and appealing side effects. This has led to officers being more verbal overall, sometimes even speaking into camera microphones to provide context such as a subject’s parole status, or if a weapon is reported before arriving at the scene. The audio transcript contains important details that DraftOne will include in its report, such as whether the
“The officer then asks further questions during the interaction,” Spitzer-Williams said. “They keep saying things like, ‘Okay, Jimmy,’ which means they agree to search your backpack.”
Spitzer-Williams said this also helps with community relations because officers are explaining what they’re doing and why.
However, an American Civil Liberties Union investigation found that police officers’ real-time narration is also being used to manipulate evidence. A common example is when an officer yells “stop resisting” to justify the use of force, even though the individual is complying. Axon’s Spitzer Williams said he doesn’t think Draft 1 will exacerbate this “real concern.”
Spitzer-Williams also pointed to an Axon study that found that reports written by software were less likely to use biased language than reports written by police officers.
Back in the East Palo Alto Police Department’s conference room, Venegas read out the AI-generated report. “On September 23, 2024, at approximately 10:49 a.m., I, Officer Venegas, responded to a call regarding a domestic disturbance involving a minor.”
She said the program produced good drafts but had some limitations. At this point, Draft One only understood English, so several things were wrong, such as who was speaking and who was related to whom.
“Sometimes you make little mistakes like that, but it’s easy to fix,” Venegas said.
Some in the criminal justice field said these seemingly small errors are indicative of larger issues with authorship that could become significant during the sentencing process. Dr. Matthew Guariglia, a policy analyst at the Electronic Frontier Foundation, said he was concerned that these reports would “destroy the ability to cross-examine police officers.” Because even if a police officer is caught lying on the stand, they can always say, “Well, the AI wrote that.” ”
If part of it is formed by AI, there are big concerns about whether we can trust AI.
andrew ferguson
White, whose Campbell Police Department has been using the software since May, said the department’s policy ensures that officers are held accountable for reports created with Draft One. He said even if an officer “miscued” and found an error in the report, it could be easily resolved by reviewing what he called the “gold standard” of evidence: body camera video. .
“Human activities are subject to error,” White said. He believes AI will make things more accurate, not less accurate.
San Mateo County Assistant District Attorney Rebecca Baum, who has been in discussions with the East Palo Alto and San Mateo Police Departments about moving to AI-assisted reporting, said her office is cautiously optimistic about the new program. said. Her main concern is that body cameras, especially body camera audio, don’t capture everything that happens during an incident.
“Body camera audio recordings don’t tell you things like witness demeanor, whether someone was under the influence of drugs or alcohol, whether there were injuries, etc.,” Boehm said.
Boehm said police officers need to be continuously involved in the process of creating these reports to ensure nothing is left out, especially information that indicates a person may be innocent. He said he had an obligation to hand it over.
Axon also has many safeguards built into the application to allow personnel to scrutinize and proofread reports to ensure accuracy.
Officer Venegas demonstrated on a computer screen how each paragraph of the report contains comments that must be resolved before the report can be moved from Draft 1 to the department system.
Then, at the bottom of the screen, Venegas pointed to a box you had to click. This was the final step, which Axon’s Spitzer-Williams said was “probably the most important” given the potential impact of each report.
“I acknowledge that this report was prepared using Axon’s Draft One,” Mr. Venegas read. “And I further acknowledge that I have reviewed the report in detail, have made any necessary edits, and believe that it is an accurate representation of my recollection of the events reported. I intend to attest to the accuracy of this report.”
This article is a collaboration between KQED, Guardian US, and California Newsroom.
