Law enforcement agencies have poured $400m into Axon’s tasers, cameras and cloud vault. AI that writes police reports is next. What could possibly go wrong? Michael Sainsbury asks.
The pitch from Axon Enterprise must sound irresistible to police commissioners everywhere. An officer presses record on a body camera, and Axon’s Draft One artificial intelligence (AI) listens to the audio and writes a complete police report, slashing paperwork and freeing officers for the beat.
Yet when prosecutors in King County, Washington, reviewed these AI-generated reports in 2024, they found a fatal flaw, outlined in a memo from a local police chief:
“In one example we have seen, an otherwise excellent report included a reference to an officer who was not even at the scene. This is one type of error that could easily go unnoticed by a reviewing officer given the volume of material required to be reviewed on deadline. And when an officer on the stand alleges that their report is accurate — they will be proven wrong.”
Police agencies were subsequently told that the court would reject any narrative written with AI assistance because of the risk of errors. In transcribing the audio, Draft One had invented a phantom officer, one of many errors emerging from the growing number of police forces using the software.
Axon, Palantir and Cellebrite
As reported by MWM, US big-data giant Palantir and Israel-founded forensics vendor Cellebrite operate inside almost every Australian security agency. The missing infrastructure bridging them is Nasdaq-listed Axon Enterprise Inc.
Public procurement commitments to Axon across Australian jurisdictions comfortably exceed $400m. Axon supplies TASERs, body cameras, dashcams and Bluetooth holsters that activate the bodycam when a gun is drawn. It also runs Axon Evidence, the cloud vault now acting as the master intake pipe for policing in Victoria, NSW, Queensland and the ACT.
Together, the three form a closed-loop surveillance triad. Axon captures field video, Cellebrite dumps extracted phone data into Axon Evidence, and Palantir ingests transcriptions and metadata to map networks and cross-reference state and federal databases.
AusTender disclosures show more than $56m in federal Axon contracts for the AFP, Border Force and other agencies. States have committed tens of millions more, including $17.8m in Western Australia and $42.6m for Victoria’s initial body camera rollout, with total Victorian spend estimated at $129m. NSW recently committed $108m to a police technology upgrade that includes Axon Evidence.
No isolated glitch
The Washington failure was no isolated glitch. Prosecutors in Oklahoma restricted AI reports in felony cases. California and Utah mandated disclosure and preservation of raw drafts. Audits in Colorado and Louisiana found bugs that let unverified AI text slip into court dockets.
Closer to home, a Law Enforcement Conduct Commission review found more than 34%, or about 2,900, of NSW Police body cameras were broken. NSW Police nevertheless ordered nearly 10,000 new Axon Body 4 units fitted with the automated holster.
Axon is now pitching Draft One to Australian jurisdictions as an upgrade to existing Axon Evidence subscriptions. It promises to automate 30% to 40% of the shift time that officers spend on paperwork.
Digital Rights Watch head of policy Tom Sulston told MWM, “The way Axon works is that it doesn’t have any change control over that report, so you can’t tell what was generated by the AI and what was actually written by a police officer. It’s the sort of thing I’d expect a good defence barrister to make absolute mincemeat of.”
Context and dialect
Sulston also raised context and dialect. Draft One runs on OpenAI’s GPT models, which were trained largely on American English. That may work so well in Australia, where police deal daily with First Nations people.
“If someone says ‘yeah’, is that just a keep-talking noise, or are they saying ‘yeah, I did that crime, you got me’?” Sulston said. “That sort of human interpretation isn’t well left to large language models because they don’t have that context”.
US critics raised the same alarm. The Electronic Frontier Foundation (EFF) warned in a 2024 analysis that the public should doubt whether a language model could reliably handle dialects, vernacular, idioms and slang. A Stanford study found commercial speech recognition systems misheard Black American speakers almost twice as often as white speakers.
Australian courts have long grappled with the problem. Linguist Diana Eades documented “gratuitous concurrence”, the tendency of Aboriginal English speakers to answer yes regardless of agreement or even understanding, especially where there is a serious power imbalance. The Federal Court accepted it as a real communicative phenomenon in 2022, Eades said in April.
A model that reads every “yeah” as agreement risks writing that pattern into a sworn police report as an admission.
AI drafting also collides with section 32 of the NSW Evidence Act 1995 and its equivalents, which govern when witnesses can use contemporaneous notes to revive their memory in the box.
“They go back to court months later, get asked ‘Is this what happened?’ and they’ll say ‘I’m not really sure, because I didn’t write it,'” Sulston said.
“It falls into the lazy thinking trap that there’s real police work, and there’s paperwork, and they’re not the same thing,” he said. “Officers having time to sit down, reflect on what they’ve just seen and write it down as an aide-memoire is valuable work. It adds value to the police, the prosecution and the courts”.
Surveillance risks
Under state surveillance laws, such as the NSW Surveillance Devices Act 2007 and the Victorian Surveillance Devices Act 1999, police body-worn video is classed as highly restricted “protected information”. Police can deploy footage to substantiate charges. Defence solicitors face onerous subpoenas or requests under the Government Information (Public Access) Act 2009 simply to view raw files.
Retention policies vary widely between jurisdictions and depend on manual police tagging. Unflagged footage can be purged within 90 days in some states, destroying evidence before defence teams or oversight bodies can check whether an AI report reflected the video.
On the other hand, unrestricted data retention holds its own dangers.
Axon’s hardware is a Trojan horse for a high-margin software-as-a-service trap. Once police upload petabytes of video to Axon Evidence, hosted on Amazon servers in Sydney and Melbourne, migrating away from the vendor becomes practically and financially unviable.
The EFF and the American Civil Liberties Union have warned that AI drafting erodes the independence of officer memory and risks laundering hallucinations into sworn legal briefs. It also leaves no audit trail. An EFF investigation found Draft One saved neither the AI draft nor any edited version, and the draft vanished once the officer closed the window.
“Why are the police allowed to have this thing when basic auditability and transparency is a core requirement of any society over its police force?” Sulston said.
Michael Sainsbury is a former China correspondent who has lived and worked across North, Southeast and South Asia for 11 years. Now based in regional Australia, he has more than 25 years’ experience writing about business, politics and human rights in Australia and the Indo-Pacific. He has worked for News Corp, Fairfax, Nikkei and a range of independent media outlets and has won multiple awards in Australia and Asia for his reporting. He is a fierce believer in the importance of independent media.