Draft One
An AI writing tool that drafts police report narratives from body-worn camera audio. Every draft is reviewed, edited, and approved by the officer before it's submitted.
:format(webp))
PRODUCT SAFEGUARDS
PRODUCT SAFEGUARDS
Human Centered
Every report requires officer review and approval
Built-In Paper Trail
Draft activity is logged in an audit trail
Tested for Fairness
Developed to promote quality and avoid bias
:format(webp))
PRODUCT SAFEGUARDS
PRODUCT SAFEGUARDS
Human Centered
Built-In Paper Trail
Tested for Fairness
Built Responsibly
This page is for community members who want to understand how Draft One works and what guardrails are in place. We believe how we build our products matters just as much as what we build. Our Responsible Innovation Framework is where that belief becomes practice.
What is Draft One
Draft One is built on a simple principle: AI should extend human capability, not replace it. Officers spend a fair amount of time on paperwork, but Draft One gives that time back, so officers can do more of what requires their presence, judgment, and expertise in the community.
The tool analyzes body-worn camera audio transcripts and officer-provided context to generate an initial police narrative report draft. Officers are required to review, edit, and finalize every draft before it becomes an official record. No report ever reaches that record without a human officer's explicit approval.
Draft One is continuously updated based on officer and community feedback, and potential bias is measured and evaluated on an ongoing basis. The sections below outline the specific safeguards we have built into Draft One, and the community input that shaped each one.
What is Draft One
Draft One is built on a simple principle: AI should extend human capability, not replace it. Officers spend a fair amount of time on paperwork, but Draft One gives that time back, so officers can do more of what requires their presence, judgment, and expertise in the community.
The tool analyzes body-worn camera audio transcripts and officer-provided context to generate an initial police narrative report draft. Officers are required to review, edit, and finalize every draft before it becomes an official record. No report ever reaches that record without a human officer's explicit approval.
Draft One is continuously updated based on officer and community feedback, and potential bias is measured and evaluated on an ongoing basis. The sections below outline the specific safeguards we have built into Draft One, and the community input that shaped each one.
:format(webp))
STEP BY STEP
How Draft One works:
STEP BY STEP
How Draft One works:
STEP BY STEP
How Draft One works:
:format(webp))
An officer responds to a call for service and conducts their typical police work. Their body-worn camera begins recording according to their agency's policies, just as it would in any other interaction.
Key safeguards
The key themes from our community research on Draft One heavily inform how we design, develop, and roll out this technology. We evaluate these insights alongside rigorous ethical and inclusion standards to shape our product development process. The resulting safeguards are a meaningful reflection of that dialogue, purposefully designed to address the core issues that surface.
:format(webp))
Human Centered
Officers must manually review and approve every report. Mandatory placeholders and optional minimum edit requirements help ensure accuracy is always human-verified before submission.
:format(webp))
Built-in Paper Trail
Draft One activity is captured in an audit trail. In the U.S. and coming soon to other countries, agencies can optionally retain AI-generated drafts for additional auditing and oversight. By default, all drafts include a disclosure that AI assistance was used, while officers retain full review and editorial responsibility for the final report.
:format(webp))
Tested for Fairness
Two racial bias studies found Draft One does not produce more negative or incriminating language based on race. Additionally, a double-blind study found Draft One produces clearer, more professional reports with no loss of neutrality or objectivity.
Human Centered
WHAT WE HEARD FROM THE COMMUNITY:
All drafts should require manual officer review and approval before submission to help ensure the reports correctly reflect the events as they happened.
IMPLEMENTED SAFEGUARDS:
Mandatory Review: Every draft requires officer review and approval before submission. Drafts use placeholder text for details that must come from the officer, information not captured in the body-worn camera audio, or when the source information is unclear or uncertain.
Optional Edit Requirement: Agencies can require a minimum percentage of edits (typically 10 to 40%) before submission is enabled
Obvious Error Mode: Agencies can enable the "intentionally insert obvious errors" setting to help ensure officers read and carefully edit every draft. You cannot submit and finalize a report that still contains obvious errors.
Deliberate Naming: "Draft One" explicitly signals that outputs are preliminary and require human review.
Human Centered
WHAT WE HEARD FROM THE COMMUNITY:
All drafts should require manual officer review and approval before submission to help ensure the reports correctly reflect the events as they happened.
IMPLEMENTED SAFEGUARDS:
Mandatory Review: Every draft requires officer review and approval before submission. Drafts use placeholder text for details that must come from the officer, information not captured in the body-worn camera audio, or when the source information is unclear or uncertain.
Optional Edit Requirement: Agencies can require a minimum percentage of edits (typically 10 to 40%) before submission is enabled
Obvious Error Mode: Agencies can enable the "intentionally insert obvious errors" setting to help ensure officers read and carefully edit every draft. You cannot submit and finalize a report that still contains obvious errors.
Deliberate Naming: "Draft One" explicitly signals that outputs are preliminary and require human review.
:format(webp))
Built-In Paper Trail
WHAT WE HEARD FROM THE COMMUNITY:
Ensuring AI-generated reports are factually accurate and officers remain accountable for their content. There should be clear disclosure when AI assistance is used in report generation, and honest communication about what the AI can and cannot do.
IMPLEMENTED SAFEGUARDS:
Built-In Audit Trail: Every time Draft One generates a narrative, that action is recorded in a digital audit trail capturing who used the tool, when, and what evidence was involved.
Original AI-generated draft retention: Draft One has an option to allow agencies to retain and access the original AI-generated narrative drafts.
Default AI Disclosure: By default, each report using Draft One includes a customizable disclosure indicating the narrative was developed using AI.
Officer Accountability: Officers retain full responsibility for report accuracy and content and must be willing to testify to the report’s accuracy. Until the placeholders are corrected, the report cannot be submitted.
Built-In Paper Trail
WHAT WE HEARD FROM THE COMMUNITY:
Ensuring AI-generated reports are factually accurate and officers remain accountable for their content. There should be clear disclosure when AI assistance is used in report generation, and honest communication about what the AI can and cannot do.
IMPLEMENTED SAFEGUARDS:
Built-In Audit Trail: Every time Draft One generates a narrative, that action is recorded in a digital audit trail capturing who used the tool, when, and what evidence was involved.
Original AI-generated draft retention: Draft One has an option to allow agencies to retain and access the original AI-generated narrative drafts.
Default AI Disclosure: By default, each report using Draft One includes a customizable disclosure indicating the narrative was developed using AI.
Officer Accountability: Officers retain full responsibility for report accuracy and content and must be willing to testify to the report’s accuracy. Until the placeholders are corrected, the report cannot be submitted.
:format(webp))
Tested for Fairness
WHAT WE HEARD FROM THE COMMUNITY:
Large language models trained on broad internet data can reflect societal biases (including those tied to race, gender, and socioeconomic status) in ways that could affect how incidents are documented.
IMPLEMENTED SAFEGUARDS:
Racial Bias Studies: Two racial bias studies found Draft One does not produce more negative or incriminating language based on race.
Quality Verified by Outside Experts: A double-blind study by 24 external experts in law enforcement, criminal law, and equity and inclusion found Draft One produces significantly better terminology and coherence than officer-only reports, with equivalent completeness, neutrality, and objectivity.
Tested for Fairness
WHAT WE HEARD FROM THE COMMUNITY:
Large language models trained on broad internet data can reflect societal biases (including those tied to race, gender, and socioeconomic status) in ways that could affect how incidents are documented.
IMPLEMENTED SAFEGUARDS:
Racial Bias Studies: Two racial bias studies found Draft One does not produce more negative or incriminating language based on race.
Quality Verified by Outside Experts: A double-blind study by 24 external experts in law enforcement, criminal law, and equity and inclusion found Draft One produces significantly better terminology and coherence than officer-only reports, with equivalent completeness, neutrality, and objectivity.
:format(webp))
FAQs
Key terms
The following definitions explain important terminology referenced throughout this page.
Body-Worn Camera
Specialized, high-definition audio-video recording devices designed for law enforcement, corrections, and security personnel to document interactions and collect digital evidence
Narrative Report
The detailed, chronological account of an incident written by a responding officer
Axon Evidence
A secure, cloud-based system that helps law enforcement agencies store, manage, and review digital evidence, including body-worn camera footage
Audit Trail
A secure, uneditable, chronological record of activities, transactions, or system changes that documents the "who, what, when, and why" behind every action to ensure the chain of custody of evidence
Bias Detection
The process of identifying, measuring, and analyzing systematic, unfair, or prejudiced patterns in data, artificial intelligence (AI) models, or decision-making systems
:format(webp))
:format(webp))