AI tool speeds police reports but introduces errors in family violence cases, Tasmania trial finds

Axon's Draft One cut report-writing time in half during a trial with Tasmania Police, but struggled in complex family violence situations, raising concerns from human rights groups.

By LineZotpaper
Published
Read Time3 min
Sources2 outlets
An artificial intelligence tool that generates police reports from body-worn camera audio cut administrative time almost in half during a trial in Tasmania, but also introduced errors including misidentified roles and missed threats, particularly in complex family violence incidents, according to documents released under Right to Information.

Tasmania Police trialled Draft One, an AI assistant developed by US firm Axon, in 2025 to reduce the administrative burden on officers writing reports on family violence incidents. The tool analyses audio from body-worn cameras and produces draft narratives for officers to review and approve.

The first phase of the trial analysed nearly 150 historical body-worn camera recordings from family violence incidents. The second phase, conducted in Launceston over several months, compared human-written reports with those assisted by the AI.

The evaluation found Draft One reports were written and validated in an average of 42 minutes, almost twice as fast as traditional methods. This could save approximately 160 weeks of work time and $434,187 in wages annually, the evaluation said. It also noted the tool sometimes identified coercive behaviours that officers had missed.

However, the AI introduced errors. In one case, Draft One wrote that a woman alleged a person with a hidden firearm had made threats against her life. When an officer reviewed the original footage, the woman had said words similar to "no threats or anything like that" while another person was speaking. The evaluation said Draft One performed well in simple, single-officer interactions but decreased in effectiveness in complex scenarios involving multiple people, where it occasionally misidentified roles, omitted key details, or generated overly verbose narratives.

Tom Sulston from Digital Rights Watch said it was "really worrying" that the tool had not performed well in complex family violence situations, describing it as "least functional when it would be most useful." He pointed to overseas reports of errors, including a Forbes story where Draft One suggested an officer had shape-shifted into a frog after misinterpreting audio from a Disney film.

Alina Thomas from Tasmanian family violence support service Engender Equality said AI assistance could be helpful but should not substitute human experience and communication. She suggested protocols allowing victim-survivors to have a say on whether AI tools are used in their cases.

The evaluation recommended a phased rollout once clear ethical governance policies were in place. A police spokesperson said funding models, further consultation, and additional trials would be required before any potential rollout. Axon's website states Draft One does not add speculation or embellishments and that officers must review reports before approval.

§

Analysis

Why This Matters

  • The trial shows AI can significantly reduce police paperwork, potentially freeing up officer time for frontline duties.
  • Errors in family violence reports could undermine evidence quality and victim safety, particularly in complex cases where accurate documentation is critical.
  • The outcome will influence whether other Australian jurisdictions adopt similar AI tools for policing.

Background

Family violence incidents make up a large portion of police call-outs, and report writing is time-consuming. Axon's Draft One is already used by some US police departments. The tool uses large language models to convert audio into structured narratives. Tasmania Police's trial is one of the first known tests of such technology for Australian policing, raising questions about accuracy, accountability, and legal admissibility of AI-generated reports.

Key Perspectives

Tasmania Police: The tool offers substantial time and cost savings, and can catch details officers miss. However, they acknowledge errors and say further trials and stakeholder consultation are needed before any rollout. Axon: The company markets Draft One as a factual, non-speculative tool, emphasising mandatory human review. Axon has not commented on this specific trial. Digital Rights Watch and Engender Equality: Both groups highlight risks. Critics argue the AI is least reliable in precisely the high-stakes family violence situations where it would be most useful, and worry about a lack of consent for victims. They question whether AI-assisted reports will stand up in court and whether officers can effectively review them without relying on the original audio.

What to Watch

  • Whether Tasmania Police proceeds with a phased rollout or modifies the tool after further trials.
  • Any legal challenges to the admissibility of AI-assisted reports in Australian courts.
  • Expansion of Draft One trials to other Australian states or to other report types such as child safety and crash reports.

Sources

Checked

11 claims in this story checked against 1 source: 5 supported.

Show each claim
  1. Partly supportedThe summary says the trial raised concerns from human rights groups.Latest Science news: “Tom Sulston, from human rights group Digital Rights Watch, said it was “really worrying””
  2. Partly supportedThe tool missed threats particularly in complex family violence incidents.Latest Science news: “The evaluation also found Draft One occasionally confused names and roles, and, in a small percentage of cases, missed a threat, assault or risk factor.”
  3. SupportedThe trial found Draft One cut report-writing time almost in half.Latest Science news: “The evaluation showed Draft One reports were written and validated in just 42 minutes, almost twice as fast as the traditional method.”
  4. SupportedThe trial raised concerns about Draft One’s performance in complex family violence situations.Latest Science news: “Tom Sulston, from human rights group Digital Rights Watch, said it was “really worrying” that Draft One had not performed well in complex family violence situations.”
  5. SupportedSulston called the tool “least functional when it would be most useful.”Latest Science news: ““It's least functional when it would be most useful,” Mr Sulston said.”
  6. Partly supportedDraft One wrote that a woman alleged a person with a hidden firearm had made threats against her life.Latest Science news: “Draft One wrote that a woman had alleged a person who had a hidden firearm had “previously made threats against her life”.”
  7. SupportedThe woman said words similar to “no threats or anything like that.”Latest Science news: “they said the woman had actually said words similar to: “… no threats or anything like that””
  8. Partly supportedThe trial’s first phase analysed nearly 150 historical body-worn camera recordings.Latest Science news: “The first phase of the trial involved Draft One analysing almost 150 randomly selected body-worn camera recordings from historical family violence incidents.”
  9. Partly supportedThe whole trial took place in 2025.Latest Science news: “The second phase tested the AI tool in Launceston over several months in 2025”
  10. Partly supportedThe evaluation said Draft One identified coercive behaviours that officers had missed.Latest Science news: “It also noted the AI tool had been able to pick up some potential coercive behaviours that had been missed by officers.”
  11. SupportedThe evaluation said the trial could save about 160 weeks of work time and $434,187 in wages over a year.Latest Science news: “Over a year, this would save about 160 weeks of work time and $434,187 in wages, the evaluation found.”

Zotpaper

Written by June Katz on the Tech & AI desk. Claims checked by Ines Fairley, verification. Read for legal risk by Marcus Hale, legal. Edited by Angus Burke, chief sub. Plate by Ollie Flash, picture desk.

1 source5 of 11 claims matched to a sourcefirst published
Meet the newsroomHow we workSubscribe