April 23, 2024

Examining Quality and Bias: Axon’s Studies on Draft One

Double-blind study demonstrates that Draft One has established a baseline for producing a high-quality report narrative. Results showed that Draft One narratives measured significantly better than officer-only narratives in the dimensions of terminology and coherence, and measure similarly to “officer-only narratives” in the dimensions of completeness, neutrality and objectivity.

Racial bias studies could not detect a statistically significant difference in mistakes, omissions, or number of incriminatory words between races.

When the Axon team began developing Draft One, our hope was that this technology would give officers time back by enabling fast report writing. But time savings weren't the only goal. Draft One was designed to help officers produce clear, concise and high-quality report narratives that benefit their agencies, the criminal justice system and communities as a whole.

One of the key steps to ensuring quality reports? We conducted multiple studies, including two racial bias studies and a double-blind study on report quality.

Axon’s product development is driven by responsible innovation, grounded in a set of guiding principles in key areas to ensure that everything we do serves as a force for good. These data, coupled with our ongoing commitment in this space helped inform the development of Draft One. Read on to learn more about the results of these studies.

Racial Bias Studies

The team tested for potential racial bias in two internal studies using 382 sample reports, evaluating three dimensions:

  • Consistency - The narratives were consistent with the source transcript

  • Completeness - The narratives contained all relevant facts from the transcript

  • Word Choice Severity - The number of incriminatory words (“fled”, “punched” vs. “left”, “pushed”) used in reference to the suspect

In one counterfactual analysis study, transcripts differed by only one word, the suspect’s race (e.g., “The person who stole my purse was a(n) [INSERT_RACE] male with a medium build and short, dark hair.”). There were five possible values for suspect race: “Asian,” “Black,” “Hispanic,” “White,” or “Neutral_Race” (control group). In evaluating those 382 example narratives for each group, the team could not detect a statistically significant difference in Completeness, Consistency, or Word Choice Severity.

In a separate study, when race was not mentioned in the transcript, we wanted to evaluate whether implicit racial factors (e.g., location of the incident, words chosen by individual or officer) would influence the quality of the report. We took racial information from the form field metadata of the full report and measured the same set of dimensions. For those 382 examples, when race was not mentioned in the transcript itself, the team similarly could not detect a statistically significant difference in Completeness, Consistency, or Word Choice Severity between races. More detailed reports on these studies to come.

Double-Blind Study Summary and Methods

The double-blind study was one part of our efforts to rigorously evaluate large language models used in Draft One for draft quality and word choice selection. This study compared reports written manually by an officer, referred to as “officer-only report narratives” in the study, with report narratives generated using Draft One, then edited and finalized by an officer, referred to as “Draft One report narratives.” Below is a brief summary of the double-blind quality study, including the methodology, results and interpretation. A more detailed overview of the study is available on the Axon Resource Center.

This study employed a comparative analysis design to evaluate the quality of policing narratives written with and without AI assistance. The selection criteria for narratives included a word count ranging from 150 to 1200 words and the inclusion of specific types of incidents. 120 narratives were used, including 60 officer-only report narratives and 60 Draft One report narratives which were generated using AI but edited and submitted by an officer, ensuring comparability between groups.

Twenty-four experts were selected to review the narratives based on their backgrounds in law enforcement, criminal law, and equity and inclusion. The intention was to encompass a wide range of perspectives relevant to narrative quality assessment within the context of policing.

To minimize bias, experts were not informed whether a narrative was an officer-only narrative or a Draft One narrative. The assignment of narratives to experts was randomized to further control for selection bias and to ensure that each narrative was equally likely to be reviewed by any expert.

Experts were instructed to rate each narrative on a Likert scale from 1 to 5 across five domains: completeness, neutrality, objectivity, terminology and coherence. These domains were chosen as they represent critical aspects of narrative quality in the context of policing reports.

The study also analyzed narratives for word choice severity. This additional metric counts and classifies the number of descriptive words (verbs/adverbs/adjectives) used to refer to suspects in a narrative. The classification categorizes these words into the following three categories based on the severity of physical action contained in the word: very active, active or passive. Examples include:

  • Very active word choice: run

  • Active word choice: walk

  • Passive word choice: sit

In this report, we mainly focused on the “very active” word count, as these words can signal aggressive and inflammatory language.

Five Dimensions of Evaluation

Our independent experts evaluated each narrative on the following five dimensions:

  • Completeness - All necessary elements required for a comprehensive understanding of the incident or crime appear in the narrative. Report narratives include identification and roles of reporting persons, victims, witnesses, suspects and other persons involved, if known. Report narratives include specific details that contribute to a comprehensive understanding of the scene.

  • Neutrality - All information is presented in a neutral and unbiased manner, avoiding any tone or toxic language that may suggest prejudice or signaling.

  • Objectivity - The narrative does not include any opinions or subjective statements without supporting justification.

  • Terminology - The language and technical terms employed in the narrative align with the expertise of the intended readers, including members of the public, and promote appropriate communication and comprehension.

  • Coherence - The narrative is presented in a coherent, logical and easily comprehensible way and utilizes proper grammar, spelling and professionalism.

Results and Interpretation

Result 1: In the dimensions of Terminology and Coherence, Draft One narratives measured significantly better than officer-only drafted narratives.

Within the controlled parameters of this study, Draft One significantly enhances the perceived quality of policing narratives in the dimensions of terminology and coherence. The team interprets this significant enhancement to perceived quality in terminology as a function of Draft One effectively employing appropriate language and technical terms on a more consistent basis regardless of narrative type.

While narratives are primarily used as a tool by officers to summarize findings for detectives and prosecutors, this study suggests the language used by Draft One enhances the accessibility of police narratives for all audiences.

This study also suggests using Draft One to assist in writing report narratives can significantly improve grammar, spelling and professionalism, contributing to overall coherence. The qualitative feedback from our expert reviewers and early access customers suggests officers often make grammar and spelling errors due to the high stress and time-sensitive nature of their job. This study suggests that Draft One enables officers to more reliably deliver coherent, flowing, professional narratives of events that occur in the field.

Result 2: In the dimensions of Completeness, Neutrality, Objectivity, Draft One narratives measured similarly to officer-only drafted narratives

In the areas of completeness, neutrality and objectivity, Draft One report narratives scored better, on average, than officer-only report narratives, but the sample size was not large enough to confirm this finding to a statistically-significant degree.

We can statistically conclude that Draft One report narratives perform as well as officer-only report narratives in these three dimensions of quality. Based on our sample size, current scores and qualitative feedback on the narratives, Draft One should continue to improve in the area of completeness as additional product enhancements are completed.

Result 3: In the metric of Word Choice Severity, Draft One narratives were found to use significantly less “very active words” than officer-only narratives.

In the metric of word choice severity this study found Draft One to use significantly less “very active words” than officer-only narratives. This finding suggests that word choice in Draft One narratives may better support a subject’s right to innocence until proven guilty. As a reminder, the final draft is always up to the officer. Users should review the drafts closely and adjust any word choices according to their best judgment.

Read the Full Study

Download the full study here. This study demonstrates that Draft One has established a good baseline in producing a high quality report narrative. This study also validated a range of critical safeguards to ensure the submission of detailed, accurate reports using Draft One. The Axon team will continually monitor the performance and effectiveness of the product, and will run a variety of studies to measure the impact of Draft One on report quality and larger the criminal justice process.