A Taxonomy of AI tools in Criminal Justice

Why Do We Need a Taxonomy?

  • To establish a common vocabulary
  • To facilitate mapping the landscape, comparing capabilities, and assessing risks

Taxonomy Framework

This taxonomy is based on:

  • The mode in which information is inferred from inputs
  • The mode of inputs and outputs

Three Modes of Inference

Analysis

Analysis of content (e.g., classify, match, cluster, transcribe)

Examples: Recognise a face or voice, identify a weapon in an image, transcribe from audio to text

Synthesis

Synthesis and enrichment (e.g., combine to predict, profile)

Examples: Profile probability to re-offend, predict spatio-temporal crime hot spots

Generation

Generation of content (e.g., generate from a prompt)

Examples: Summarisation of a statement, generation of questions for a witness interview, RAG (retrieval augmented generation)

Seven Modes of Input and Output

Primary Data

  • text
  • image
  • video
  • audio

Structured or Enumerated Data

  • enum - Enumerated types (e.g., Yes/No, sentiments, semantic features, points to prove)

Instructions for Generative AI

  • prompt - Set of instructions that impacts the result

Formal Syntax

A category in the taxonomy has the form:

Inference(Input_Modes → Output_Modes)

Where:

  • Inference = Analysis | Synthesis | Generation
  • Modes can be individual items, products (separated by commas), or lists (in brackets)

Example Categorizations

Analysis Examples

Category Purpose
Analysis(audio → text) Audio to text transcription
Analysis(audio, audio → enum) Speaker recognition (voice recognition)
Analysis(image → enum) Identification of weapon in a scene
Analysis(text → text) Translation from Welsh to English

Synthesis Examples

Category Purpose
Synthesis(text → enum) Synthesis of Points to prove from witness statement
Synthesis(enum, enum → enum) Crime hot spot prediction

Generation Examples

Category Purpose
Generation(prompt, text → text) Summarise this text
Generation(prompt, text → [text]) Generate a list of interview questions
Generation(prompt, text → image) Generate a suspect image from witness statement

Criminal Justice Stages in England & Wales

Tools are categorized by the stage(s) of the criminal justice process where they are deployed. Based on the paradigm of England & Wales, we split the criminal justice system in the following stages:

Stage Name Purpose
1 Community Policing and Offender Management Crime prevention, public protection, identifying vulnerable individuals
2 Intelligence Identify offenders and wider networks
3 Investigation Gathering evidence, pursuing lines of inquiry, victim protection
4 Charging Decision or Alternative Disposal Decision on charging or alternative disposal
5 Trial or Guilty Plea Defendant admits guilt or opts for trial
6 Sentencing Appropriate sentence decided
7 Prison and Parole Management of offenders in custody
8 Probation Management of released offenders

Data Structure

Each AI tool in our database contains the following fields:

  • Tool Name - The full name of the tool
  • Purpose - A concise statement of the tool's intended purpose
  • Description - Fuller explanation of operation, scope, and functionality
  • Users - Organizations deploying or trialling the tool
  • Deployment Stage - Current stage (Trialled, Deployed, Experimental, Stage unknown)
  • Development Type - In-house, Third-party, or Academic collaboration
  • Developer/Vendor - Entity responsible for developing the tool
  • Criminal Justice Stages - Stages of the process where the tool applies
  • Taxonomy - Classification using our taxonomy framework
  • Tool Functionality - Brief description corresponding to taxonomy
  • Resources - Links to publicly available sources