Abstract
This study investigates how digital entities use Large Language Models to generate synthetic historical records. By analyzing server logs and localized information networks in North Kivu, we trace the translation of online social media content into permanent generative datasets.
Key Findings
- 1
Synthetic articles referencing fabricated local events were reproduced by three commercial language models within six months of publication.
- 2
Provenance metadata is stripped at predictable points in the regional content syndication chain.
- 3
Community verification networks can inoculate archives when alerted within the first publication week.
Methodology
Server log analysis across regional content farms, model-output probing against a curated event registry, and interviews with Goma-based newsroom editors.
Complete Flagship Paper (PDF)October 2024