Selective Disclosure

How the Police Shape the Supply of Crime Information

Ashrakat Elshehawy

University College London

Violeta Haas

Trinity College Dublin

Sascha Riaz

Singapore Management University

Crime rates are decreasing…

Property/theft: DE PKS Diebstahl insgesamt / US FBI property crime. Violent: DE PKS Gewaltkriminalität / US FBI violent crime.

…yet worries about crime are increasing

Germany United States
Roose 2021 Gallup 2023
43% feel unsafe at night 40%
66% view crime as a (very) important / serious issue 63%
62% believe crime is increasing 77%

What, then, shapes public concern about crime?

Traditional answer, traditional actor

Direct experience rare → concern tracks information exposure, not objective risk (Hale 1996; Esberg & Mummolo 2018)

News media:

  • Relative to police records, news overrepresents
    • Rare violent offenses (Ihle 2015; van Um 2015)
    • Foreign/minority suspects (Dixon & Linz 2000; Arendt et al. 2017; Hestermann 2019, 2022)
  • Exposure shapes perceptions and behavior
    • Voters link crime to out-groups + worry rises with immigration (Homola 2021; Ajzenman et al. 2023)
    • Out-group-attributed crimes trigger hate crime + raises anti-immigrant voting (Quillian & Pager 2001; Frey 2020; Couttenier et al. 2021; Riaz et al. 2024)

This study

Blind spot!
Recorded incident
No Police Press Release
Police Press Release
Media coverage of crime
Public perceptions

Police preselect what newsrooms see!

Two margins of discretion:

  1. Which incidents to publicize
  2. What details to include

→ First systematic study measuring these selection steps at scale.

Observable Context: Germany

  • Organizational structure:
    • Policing is decentralized to the 16 states
    • Overseen by a state’s Interior Ministry
  • Press releases:
    • Written by small press offices embedded in local police stations
    • Usually staffed by street-level officers
  • Reporting rules:
    • State press laws & internal directives
    • German Press Code §12.1

“As a rule, group membership should not be mentioned, unless there is a justified public interest.” (Pressekodex, Richtlinie 12.1)

Corpus construction, 2014–2025

Example: Press Release

Why an open-weight LLM

  • Why not classic NLP
    • Releases are written by hundreds of different humans, each with their own prose and linguistic conventions
    • Out-group cues are heterogeneous and contextual (nationality, appearance, accent, language)
    • Keyword/dictionary methods need stable lexicons and miss that ambiguity (Grimmer & Stewart 2013, Nelson et al. 2021)
    • Supervised classifiers need large hand-labeled data that scale poorly to many rare categories (Wilkerson & Casas 2017, Laurer et al. 2024)

→ LLMs fit this classification setting better (Gilardi et al. 2023)

  • Why open weights, self-hosted
    • Commercial APIs change and are expensive (GPT-5.5 Batch API ~$95,000)

→ DeepSeek-V4-Flash is archivable, reproducible, free and competitive with commercial APIs on political-science classification (Hilbig 2026, Spirling 2023, Higton et al. 2026)

Scale of task

  • 2.65 million press releases, 2014–2025

    2.65m releases × 3 minutes ÷ ~2,080 hours/year ≈ 64 years of human coding

  • 3.89 million events (1.5 events per release)

  • Model weights alone ~152 GB (DeepSeek-V4-Flash)

16.2 million calls on U.K.’s fastest and most powerful (11th in the world) AI supercomputer

Classification Pipeline