Can Prompted LLMs Fake Benford?

Event: DySES 2026 - Dynamics of Socio-Economic Systems
Location: Naples, Italy
Date: June 18-21, 2026

This talk investigates whether large language models can generate synthetic accounting figures that resemble Benford-compliant numerical data. Using a controlled prompt-level experiment, I compare four conditions: legitimate accounting amounts, fraudulent accounting amounts, legitimate amounts generated with explicit Benford guidance, and fraudulent amounts generated with explicit Benford guidance.

The central question is whether first-digit tests can detect differences between ordinary LLM-generated numerical data and data produced under prompts that explicitly ask the model to respect Benford’s Law. The talk therefore connects Benford diagnostics, synthetic data generation, prompt engineering, and fraud analytics, showing how the wording of a prompt can affect not only the explanation produced by an LLM, but also the numerical distribution it generates.