Research Summary: Using Codebooks to Detect Cybercrime Topics in Text Narratives

Increasing cybercrime-related consumer complaints are burdening state and city governments in the United States, exacerbated by reduced federal agency staffing. A novel method proposes leveraging Large Language Models (LLMs) with qualitative cybercrime research codebooks to detect specific cybercrime topics, such as impostor scams and identity theft, in consumer narratives. This approach has demonstrated high precision and recall across various LLM families, offering a scalable solution for local governments that lack resources for specialized AI model development.

Why it matters

This development is strategically important as it addresses a growing public safety and resource allocation challenge for sub-national governmental entities. It offers a practical and accessible AI-driven solution to enhance cybercrime detection capabilities, which can lead to more efficient response mechanisms and improved consumer protection. The method’s reliance on existing LLMs and qualitative research frameworks could democratize advanced analytical tools for organizations with limited dedicated technology budgets.

Key insights

• Cybercrime complaint management is shifting towards state and city governments in the U.S. due to federal agency de-staffing.
• Traditional specialized AI models for cybercrime detection are resource-intensive, posing challenges for local government adoption.
• A new LLM prompting method uses qualitative cybercrime research codebooks to detect relevant topics in text complaints.
• The method achieved high precision and recall for detecting impostor scams and identity theft across Gemini and GPT model families.
• This strategy offers a potential path for local governments to utilize advanced AI for cybercrime detection without developing custom models.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.16000