AI-generated police reports are moving from experiment to operational tool, and the central dispute is no longer only whether the technology can write a usable narrative. The bigger question is whether anyone outside the process can later understand what the AI produced, what an officer changed, and whether the final report accurately reflects the underlying body camera audio.
A new investigation from the Electronic Frontier Foundation focuses on Axon’s Draft One, a tool that uses a ChatGPT variant to create report narratives from body camera audio. The group argues that the product’s design makes meaningful oversight unusually hard, while Axon says the system includes safeguards and follows the existing police practice of saving only the final approved report.
Why Draft One Is Drawing Attention
Draft One debuted last summer at a police department in Colorado. Its basic workflow is straightforward: the tool generates an initial police report from body-worn camera audio, then officers are expected to review and edit the output.
That review step matters because the source material is not a simple form field. Police reports can shape investigations, court proceedings, and public understanding of an encounter. If AI introduces an error, misunderstands slang, adds context that was not present, or frames an event with biased language, the officer’s edit is supposed to catch and correct it.
The EFF’s concern is that the system does not preserve enough information to show whether that review happened in a meaningful way. According to the group, Draft One does not save drafts or keep a record identifying which parts of a report came from AI. Departments also do not retain multiple versions of drafts, which makes it difficult to compare the machine-generated narrative with the final text.
In the EFF’s view, that missing record blocks a basic accountability question: whether the technology is accurate enough to rely on in police reporting.
The Audit Trail Dispute
The investigation argues that Draft One “seems designed to stymie any attempts at auditing, transparency, and accountability.” The EFF says that in every department, officers do not necessarily have to disclose when AI was used, and the absence of saved drafts makes later review much harder.
The group also says it is difficult to know whether officers are carefully editing AI-generated reports or “reflexively rubber-stamping the drafts to move on as quickly as possible.” That concern is heightened by Axon’s disclosure to at least one police department that engineers found a bug that allowed officers on at least three occasions to bypass guardrails intended to prevent submission without reading first.
The EFF sees a broader legal risk. If a report contains biased language, inaccuracies, misinterpretations, or lies, the missing draft history can make responsibility harder to assign. The group warned that AI reports can create a “smokescreen” for officers, because there may be no preserved record showing whether a disputed passage came from the officer or the tool.
“There’s no record showing whether the culprit was the officer or the AI,” the EFF said. “This makes it extremely difficult if not impossible to assess how the system affects justice outcomes over time.”
The group also reviewed a video from a roundtable discussion in which an Axon senior principal product manager for generative AI described disappearing drafts as intentional. The manager said, “we don’t store the original draft and that’s by design and that’s really because the last thing we want to do is create more disclosure headaches for our customers and our attorney’s offices.”
Axon’s Response
Axon disputes the idea that Draft One removes responsibility from officers. In a statement to Ars, an Axon spokesperson said the tool drafts an initial narrative strictly from the audio transcript of the body-worn camera recording and includes safeguards, including human decision-making at crucial points and transparency about use.
Axon said officers remain fully responsible for the content of every report. According to the company, each report must be edited, reviewed, and approved by a human officer before it becomes final.
The company also said Draft One was designed to mirror the existing police narrative process, where only the final approved report is saved and discoverable, rather than interim edits, additions, or deletions made during officer or supervisor review.
Axon further said that since day one, whenever Draft One generates an initial narrative, that use is stored in Axon Evidence’s unalterable digital audit trail and can be retrieved by agencies on any report. The company said Draft One reports include a customizable disclaimer by default, and that it recently added the ability for agencies to export Draft One usage reports showing how many drafts have been generated and submitted per user, as well as reports on which evidence items were used with Draft One.
What Oversight Still Cannot Easily See
The EFF says those features do not solve the core oversight problem. In a press release, the group said its investigation “found the product offers meager oversight features,” including an audit log function it considered practically useless for broader accountability.
One key issue is scale. The EFF said Axon’s tool does not allow departments to export a list of all police officers who have used Draft One, or export a list of all reports created by Draft One unless the department has customized its process. Instead, the EFF said Axon allows exports of basic logs tied to a particular report or an individual user’s activity, such as logins and uploads.
That makes routine analysis difficult. To understand how often officers use the technology, a reviewer may have to search individual records. The EFF said that could require combing through dozens, hundreds, or in some cases, thousands of individual user logs. A similar burden applies to reviewing reports one by one because of the sheer number of reports generated by any given agency.
- Saved AI drafts could show what the tool originally wrote.
- Version histories could show what an officer changed before approval.
- Department-wide exports could show how widely Draft One is being used.
- Clear disclosure practices could help courts, defense attorneys, journalists, and civil liberties groups identify AI-assisted reports.
The EFF is calling for a nationwide effort to monitor AI-generated police reports and published a guide to help journalists and others submit records requests. But the group’s own investigation found that obtaining those records is often difficult, and in many cases, impossible.
The Stakes For AI In Police Reporting
The dispute over Draft One shows why AI in police work cannot be evaluated only by speed or convenience. A report-writing tool may reduce time spent drafting narratives, but the value of that efficiency depends on whether the final document can still be tested, challenged, and understood.
Axon’s position is that human officers remain accountable and that Draft One fits within the existing police reporting process. The EFF’s position is that AI changes the nature of that process because the machine-generated starting point may affect the final narrative, yet disappear before outsiders can inspect it.
That tension is likely to define the next phase of AI-generated police reports. If departments use the technology, the public accountability question will not be limited to whether AI was involved. It will also be whether the evidence trail is strong enough to show what happened after the tool entered the reporting process.