Responsible Innovation
Draft One
An AI writing tool that drafts police reports from body-worn camera audio. Every report 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
Narrative draft activities are logged in an audit trail
Tested for Fairness
Conducted studies to evaluate quality and racial 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, what guardrails are in place, and why we made the choices we did. 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 belief: AI should extend human capability, not replace it. 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.
The goal is to give officers back time spent on paperwork so they can do more of what requires their presence, judgment, and expertise in the community. 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 belief: AI should extend human capability, not replace it. 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.
The goal is to give officers back time spent on paperwork so they can do more of what requires their presence, judgment, and expertise in the community. 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))
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
Every AI-generated draft’s activities are logged in an unalterable audit trail. By default reports include a disclosure that AI assistance was used. Officers hold full editorial responsibility.
:format(webp))
Tested for Fairness
Two racial bias studies found Draft One does not produce more negative or incriminating language based on race. 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 reports 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: Reports require manual officer review and approval before submission. Drafts include INSERT placeholders that prompt officers to add their own observations and details not captured in the audio
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 ensure officers read and carefully edit every draft
Deliberate Naming: "Draft One" explicitly signals that outputs are preliminary and require human review
Human Centered
WHAT WE HEARD FROM THE COMMUNITY:
All reports 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: Reports require manual officer review and approval before submission. Drafts include INSERT placeholders that prompt officers to add their own observations and details not captured in the audio
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 ensure officers read and carefully edit every draft
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. An officer should control and have full editorial responsibility over AI-suggested 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:
Permanent Audit Trail: Every time Draft One generates a narrative, that action is recorded in a permanent digital audit trail capturing who used the tool, when, and what evidence was involved
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 flagged errors 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. An officer should control and have full editorial responsibility over AI-suggested 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:
Permanent Audit Trail: Every time Draft One generates a narrative, that action is recorded in a permanent digital audit trail capturing who used the tool, when, and what evidence was involved
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 flagged errors 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))
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 or conducts their typical field operations. Their body-worn camera starts recording according to their agency's policies, just as it would in any other interaction.
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
Natural Language Processing
A branch of artificial intelligence (AI) and computer science that enables computers to understand, interpret, manipulate, and generate human language (text or speech)
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))