Artificial Intelligence (AI) answer engines are fielding a growing share of inquiries from analysts, scholars, and the public regarding issues of peace and conflict. However, a new study published on arXiv warns that Large Language Models (LLMs) exhibit discernible error patterns when processing conflict-related data.
Research Methodology
Researchers posed a battery of questions concerning 28 global conflicts to five leading AI answer engines. A total of 5,460 answers were scored against documented evidence to determine the reliability of AI in geopolitical analysis. The findings reveal a direct link between the volume of retrievable data and the accuracy of AI outputs.
The 'Thin Record' Vulnerability
The study found that the thinner the retrievable record surrounding a given conflict, the more likely AI engines are to invent, misattribute, and miscount facts. These 'thin records' do not merely encourage hallucinations; they create a structural exposure to mis- and disinformation.
Because these records are sparse, they are the easiest to warp through Generative Engine Optimization (GEO). This technique allows actors to bias engine responses by optimizing the sources that LLMs crawl for information.
GEO Information Warfare
Through an analysis of 1,048 websites utilized by LLMs for conflict facts, the study confirmed that GEO source optimization is already occurring. While state-partisan digital capture remains in its incipient stages, the researchers characterize its growth as rapid.
The findings suggest a critical shift for scholarship and policy. The researchers argue for a return to deep local monitoring and translation-based research—methodologies that AI tools currently cannot replicate—to counter the rise of GEO-based information warfare.
