AI Policy & Governance, Government Surveillance
Automated Police Report Drafting
Automated police report drafting tools are increasingly advertised as shortcuts to help officers speed through a “tedious” task. But just because work is routine does not make it unimportant. Relying on AI to draft police reports carries serious risks that require significant guardrails to protect our civil rights.
Automated police report drafting refers to software systems that use artificial intelligence (AI) — often speech-to-text processing and large language models (LLMs) — to generate first drafts of police incident reports. These reports, sometimes called narratives, are official law enforcement documents that are intended to record an officer’s immediate recollection of the factual circumstances of an encounter, the actions taken, and the observed behavior of those involved. The form and substance of police reports can materially affect criminal and civil legal outcomes, especially in low-level criminal cases where the police report will often form the primary basis for charging decisions, pretrial detention, plea bargains, and sentencing.
Typical products in this space, such as Axon’s Draft One and Truleo’s AI Assistant, take audio or video recordings from body-worn cameras (BWC) and automatically transcribe them. Some of these products claim to support transcription and translation of Spanish, but others only support English. The systems then use generative AI to create a structured narrative draft of a police report that an officer can review, edit, and finalize. Other tools, like CopEntry, PoliceNarratives.ai, and Truleo’s Field Notes, take structured field notes or call data without BWC recordings and similarly transform them into narrative text. These tools promise to free up police time, and could also make it easier to produce a report sooner after an encounter, when important details that might later be forgotten are still top-of-mind. They could also make it more likely that a police report is written at all.
In general, automated report drafting tools are designed to prompt officers to provide additional information, as well as to review and edit the text drafted by the AI system. But in practice, it can be difficult if not impossible to determine how much human review and revision actually occurred before the report was finalized. Some tools include auditing features that track changes made to the initial AI-generated draft, but others only log the initial draft request and officer sign-off. In the latter circumstance, no one reviewing a police report, including judges, defendants, and even prosecutors, will be able to independently determine what information came from the officer, and what was generated by AI. Remarkably, some vendors even advertise these more limited audit trails as a feature to “reduce FOIA exposure,” or in other words: avoid accountability.
RISKS
How it might not work…
Police reports are inherently one-sided accounts of often tense, fast-moving encounters, drafted solely from the officer’s perspective. Civilians generally have no right to review or contest a draft before it becomes the official record, and once a report is finalized, officers, prosecutors, and courts often treat it as a fixed account of what happened. Because police reports often shape charging decisions, pretrial detention, plea negotiations, and trial outcomes, even small errors or distortions can materially affect whether a person is deprived of their liberty or convicted.
This makes the use of automated report drafting tools especially concerning because AI systems are prone to errors and hallucinations. When relying on speech-to-text transcriptions, especially from noisy or low-quality BWC audio, transcription errors are likely. AI systems may not handle accents, overlapping speech, or background noise reliably. Even in ideal environments, specialized speech recognition models exhibit word error rates around 10%, meaning 1-in-10 words are mistranscribed. These errors can compound when fed into generative models, as LLMs can unpredictably misinterpret context, misidentify speakers, or hallucinate unverified details into draft reports. If an AI mishears a command or hallucinates a suspect’s statement, it may produce an inaccurate narrative that could go unnoticed and result in a wrongful arrest.
Another significant risk is cognitive offloading and automation bias: Officers may rely on AI drafts rather than their own careful recollection and judgment about what happened. This could diminish memory retention and factual scrutiny, and lead to superficial acceptance of text that does not accurately reflect the encounter. Alarmingly, marketing materials for many automated police report drafting tools highlight cognitive offloading as a feature, not a failure.
These risks have led to some truly bizarre errors, including one instance where Axon’s Draft One claimed an officer transformed into a frog because of a movie playing in the background of a recording. But errors will often not be so easy to catch. More subtle mistakes, like misattributing a statement to the wrong speaker, could be more difficult to identify but have far more serious consequences.
Even if it works…
Even if an AI system generates drafts that closely match the evidence to which it has access, serious civil rights and civil liberties concerns persist:
- Judicial reliance, due process, and transparency: When AI-drafted police reports are treated as neutral or reliable accounts, they can have a major impact on charging decisions, bail determinations, and plea negotiations before any meaningful adversarial testing occurs. If the use of AI is undisclosed, or audit trails are unavailable, judges, defendants, and even prosecutors may never learn that an AI system generated key factual assertions. This deprives all accused persons of a meaningful opportunity to challenge accuracy, bias, or reliability, and undermines the transparency of the entire judicial process.
- Replacing accountability and explanation with automation: Broad adoption of AI could change the role of police reports from officer-written attestations based on personal recollections to machine-generated dramatizations that could lack critical context that is not captured by the mere transcription of an encounter, like prior interactions with an individual or activity outside the range of the recording device. Police reports are meant to capture an officer’s perspective and decision-making process. Diluting officer engagement in drafting undermines that purpose, and undermines the ability to scrutinize whether an officer’s actions were proper.
KNOWN USES
Deployment of automated drafting tools is underway in many U.S. jurisdictions. For instance, Axon’s Draft One has already been extensively piloted or deployed by police departments in Minnesota, Colorado, Indiana, and California. Meanwhile, Truleo claims to be “trusted by 1000+ departments” nationwide.
RECOMMENDATIONS
Policy frameworks should address both operational use and civil rights safeguards:
- Procurement standards: Agencies should adopt policies requiring vendor transparency around system design and known limitations, and independent verification of model and system accuracy (especially for transcription of BWC) prior to procurement.
- Officer training and accountability: Agencies should require officer training specific to automated report drafting tools, focused on identifying errors such as transcription inaccuracies, speaker misattribution, and hallucinated details. Training should be paired with certification, periodic re-training, and performance auditing.
- Responsible use policies: Agencies should adopt policies restricting use of AI-drafted police reports to low-stakes encounters (e.g., non-violent civil citations) until robust empirical validation is available, in order to prevent premature application in high-stakes contexts. Responsible use policies should also require rigorous comparison between AI-generated drafts and source evidence prior to officer approval.
- Mandatory transparency and audit trails: Laws should require that any use of AI to draft police reports is clearly disclosed in the final document and that original AI drafts and edit histories are retained for the full duration of the case record. California’s SB 524 enacted such requirements, mandating labeling of AI-assisted content and preservation of drafts. Utah’s SB 180 similarly requires disclosure of the use of AI, but fails to require full preservation.
- Discovery and defense rights: Rules for disclosure in criminal legal proceedings — including state and federal rules of criminal procedure, prosecutorial disclosure policies, and judicial discovery orders in individual cases — should explicitly encompass the use of AI. Rules should require defense access to audit trails, including original AI drafts and edit histories, and model explanations where applicable to support meaningful challenges to AI-generated evidence.
- Bias mitigation and oversight: Independent auditing of AI training data, model behavior, and outputs — particularly for differential impacts on protected groups — is necessary to detect and mitigate systemic bias. Public reporting and community oversight mechanisms can build accountability.
CONCLUSION
Automated police report drafting tools may promise to save time, but they operate at a critical juncture in policing when officers decide who to believe, and who to arrest. Without strong guardrails like required disclosures, audit trails, and limits on use, these tools risk creating unnoticed errors, automating bias, obscuring accountability, and undermining justice.