Guardrail or Guide? Generative AI, Criminal Capital, and the Re-framing of Downstream Search for Illicit Drugs

Cybercrime requires specialized knowledge, vocabulary, and an understanding of how the different parts of a task fit together. Criminal capital theory has traditionally explained how people acquire such knowledge through experience, mentors, and criminal networks. We argue that Generative AI (GenAI) introduces a new way of accessing this knowledge. Unlike conventional search engines, which require users to know what keywords to search for, GenAI can help users translate an unclear goal into more specific sub-questions in the required vocabulary. We examine whether access to ChatGPT changes how high-risk drug-related information searches are subsequently expressed on downstream conventional search engines. We exploit ChatGPT’s geographic availability following its late-2022 release and analyze 4,720,859 Google Search Console query observations across 2,965 domains and six markets from January 2021 to June 2024. Estimating difference-in-differences effects in four-week event-time bins with block-bootstrap inference, we report that for illicit-substance queries, ChatGPT accessibility is associated with lower impressions, higher click-through rates, longer queries, and greater lexical specificity. We extend criminal capital theory by showing how conversational AI may support access to crime-relevant knowledge. We also show why AI guardrails should be evaluated across platforms rather than only at the point where a response is refused, as the guardrails usually block direct execution locally but may leave informational scaffolds for continued search elsewhere.

Working Paper — Runner-Up, Best Paper Award, CSWIM 2026 (Hainan)