Suspect Freed Over AI Police Report: What Draft One and Truleo Really Deliver

What happened

On October 29, 2026, an Arizona Department of Public Safety (DPS) trooper clocked a car doing more than 100 mph on I-10 near Phoenix, where the limit was 65. The driver, Anthony Edward Rene Quidera, was arrested for driving under the influence (DUI), speeding and having an open container in the car. According to KGUN, the officer's report described the car weaving, a smell of alcohol, bloodshot eyes and six out of six clues on the eye-movement test. Quidera reportedly refused a breath test.

Commissioner Daniel Fenzel of Maricopa County Superior Court released him anyway. At the end of the probable cause statement was a note in capital letters: the report had been “created with the assistance of Truleo AI,” and its content reflected the officer's own observations. According to ABC15, Fenzel wrote that he couldn't judge whether the factual claims had been written wholly or partly by generative AI, because the officer's original account wasn't available. His conclusion:

“The court is currently unable to find probable cause.”

As far as we know, this is the first court decision to reject an AI-assisted police report as the basis for probable cause. That doesn't close the case. According to Justice Court records, Quidera was later charged, with an arraignment set for November 12. We found no statement from DPS or prosecutors.

The case raises a bigger question. Do tools like Axon's Draft One and Truleo actually save police time? Or do they just move the work into the courtroom, where judges and defense lawyers have to work out who wrote which sentence?

What the vendors promise

Draft One turns the audio from a body camera, the camera officers wear on their uniform, into a draft report. When it launched in April 2024, Axon promised the tool would save officers “an hour or more per day.” One figure from Fort Collins, Colorado, is still quoted today: 82 percent less writing time. In a later case study with Microsoft, Axon itself calls that a peak value and says most officers cut their writing time in half. According to Proof News, the figure came from officers' self-reports. Nobody measured it.

We checked the other vendor numbers. According to Proof News, Draft One saved 2.2 million minutes across 100,000 reports in its first year, roughly 22 minutes per report. Forbes Switzerland quotes Axon as saying it saved about 300,000 hours across more than 600,000 reports, roughly 30 minutes per report. By Axon's own account, that number is based on customer surveys. Competitors are more cautious too. In August 2026, Caliber Public Safety advertised savings of “up to 40 percent,” but the press release itself only mentions “expected” savings of 20 to 40 percent during the pilot phase. Truleo has been credited with “up to 50 percent,” but we couldn't find a source for it.

Truleo is an awkward case, since it's the tool from the Arizona incident. According to NBC, Truleo CEO Anthony Tassone said of body-cam-based rivals: “Studies show it doesn't save officers time.” His product, “Field Notes,” works differently, he says: it uses voice notes the officer records at the scene. According to Criminal Legal News, Tassone also warned that if a report is just an AI rewording of the body-cam video, you no longer have two independent pieces of evidence. You have one, plus a summary derived from it. That's exactly the ambiguity Fenzel objected to.

What independent measurements show

The most important study is the randomized trial by Ian Adams and Lieutenant Matthew Barter. Randomized means chance decided which reports were written with AI and which without, just like in a drug trial. Because the study was preregistered, its methods and analysis were fixed in advance. The result: writing with AI was no faster. According to Forbes Switzerland, the Manchester police stopped using it afterward. In Anchorage, Deputy Chief Gina Burington told Alaska Public Media that the time spent reviewing and correcting drafts ate up the savings. There was also a more basic problem:

“If they saw something but didn't say it, of course the body cam doesn't know that.”

The follow-up study is even more troubling. According to the ACLU, it used 92 experienced supervisors who had been signing off on reports for an average of 22 years. The AI texts were harder to read, overall quality was no better, and on accuracy they dropped on average from the 50th to the 36th percentile. Reviewers spotted the AI reports no better than they would have by guessing. One caveat: we could only check this work through secondary sources, and it's a preprint, so it hasn't been peer reviewed yet. Axon points to its own double-blind study, which found Draft One performs at least as well.

There are positive voices too. Arvada, Colorado, reports 50 to 60 percent higher efficiency. Scottsdale says it saves at least 30 minutes per report. Heber City, Utah, talks about “six to eight hours.” These figures come from the departments themselves, not from measurements with a control group.

Frog cop and phantom officer: the known errors

One case from Heber City, Utah, made headlines worldwide in January 2026. Draft One wrote in a report that an officer had turned into a frog. Disney's “The Princess and the Frog” was playing on a TV in the background, and the AI picked up dialogue from the movie. A case in King County, Washington, was more serious: a report named an officer who was never at the scene. In September 2024, the prosecutor's office said it would no longer accept AI-assisted reports. In Lafayette, Indiana, officers reported wrong names and license plates, according to Forbes Switzerland.

Then there's a structural problem uncovered by the EFF. Draft One doesn't save the first draft. Once the officer copies the text into the records system and closes the window, the draft is gone. Axon's response: “That's by design.” As a result, it's often impossible to tell which sentence came from the AI and which from the officer. According to the EFF, there was also a bug that let officers skip the requirement to read the draft, and it happened at least three times. Legal scholar Andrew Ferguson calls this “generative suspicion”: AI-generated suspicion replaces the officer's verified observation in the probable cause statement. He advises defense lawyers to request the prompts, meaning the instructions given to the AI, along with drafts and edit logs.

A September 2026 case study by the Council on Criminal Justice works through a deployment at a fictional department with 200 officers. It rates the use as a “substantial risk” and recommends only a “cautious rollout.” Its key recommendation fits Arizona almost word for word: officers should record their own recollection before reading the AI draft. The system prompt, first draft and final report should all be disclosed together as evidence. The study also warns about automation bias, the tendency to trust a smoothly written text instead of your own memory. And it notes that the strongest independent evidence, the randomized study, contradicts the vendors' time-saving claims.

What the law requires

Utah became the first state with rules in 2025 (SB 180, in effect since May 7, 2025). Every agency needs an AI policy, AI-assisted reports must be labeled, and the officer must confirm they've checked the report for accuracy. We couldn't confirm any requirement in Utah to keep the first draft.

California goes further (SB 524, Section 13663 of the Penal Code, in effect since January 1, 2026). The label has to appear on every page, the first AI draft must be kept as long as the final report, and drafts don't count as the officer's statement. Vendors may not sell or share police data. According to the EFF, Draft One can't meet this retention requirement out of the box.

Arizona has no such rule. The Arizona Supreme Court has asked its AI committee to propose rules by March 1, 2027, but those cover AI use by the courts. Meanwhile, the Phoenix police are testing Draft One until December 31 as part of a free 90-day trial.

EU: Contrary to what's often claimed, AI report drafts aren't explicitly on the AI Act's high-risk list. For law enforcement, Annex III point 6 lists things like polygraphs, recidivism predictions, profiling and systems that “evaluate the reliability of evidence.” Whether a draft report built from body-cam audio falls under that is an open question. On top of that, the “Digital Omnibus” (Regulation 2026/1744) pushed the obligations for Annex III systems back from August 2, 2026, to December 2, 2027.

What about German-speaking Europe?

According to Netzwoche, the Police Région Morges in the canton of Vaud (about 50 officers) has been using the tool “Trinity” from Deeplink since spring 2026. It doesn't work from body-cam audio. Instead it uses the minute-by-minute data logged by the dispatch center, plus voice notes and photos. It runs on an isolated network. Officers review and approve every draft. The goal was 10 percent savings. By one internal estimate, a report that used to take 30 to 45 minutes is now done “at least twice as fast.” That hasn't been measured. Switzerland isn't directly covered by the AI Act. We found no comparable pilot in Germany or Austria. According to profil, Austria's police union sees potential relief “in documentation and administration,” but no concrete trial is known.

Our take

The marketing figures of 82 or 50 percent rest on self-reports. The only measurements with a control group find no time savings, or none that's significant. The fact that officers still feel faster is a finding in itself, but it isn't proof. What is proven: errors happen, reviewers barely catch them, and without extra rules Draft One often leaves no way to show who wrote what.

The Quidera case shows where that leads. The work supposedly saved on writing comes back as an evidence question in court. A disclosure label alone doesn't fix that. In Arizona, it's what exposed the problem in the first place. What would help is keeping the first draft, as California requires, and having officers write their own note before seeing the AI draft, as the CCJ recommends.

How much weight the decision carries is still unclear. A commissioner at an initial appearance doesn't set precedent, and the suspect was charged later anyway. For police departments, it's still a warning: if you can't show what the officer actually saw, a case can fall apart at the first hurdle.