136-Country, 118M-Article Dataset Audits State Media Share in LLM Training Corpora
deliprao · x · 2026-10-22
Researchers from Google, UPenn, Northwestern and others released WorldStateMedia, an open dataset of 118 million state media articles from 136 countries in 114 languages, extracted from existing pretraining corpora with each outlet annotated by degree of state control.
Key findings:
- In English web crawls, pages from editorially controlled (vs. independent) state outlets rose from 39% in 2013 to 93% in 2025
- State media make up a substantial share of web data in some low-resource languages: 17.8% for Uyghur, 7.8% for Belarusian
- The dataset enables auditing how state narratives embedded in training and retrieval corpora shape LLM outputs
More from Safety
- Dev jokes that their SSL root cert's private key is just '5' — kuza55 · 2026-10-22
- GitLab hit by CVSS 10.0 unauthenticated RCE, patches shipped for releases back to 2014 — jd_pressman · 2026-10-22
- PHP Foundation timeline shows Calif, OpenAI, and Wiz all reported the same RCE — basedjensen · 2026-10-22
- Researcher warns GitLab flaw is 'really bad,' urges users to patch now — basedjensen · 2026-10-22
- RCE free-for-all feels like 2000 again, vet says, pointing to PHP2Shell race bugs — moyix · 2026-10-22
- Microsoft: Storm-0940 used agentic AI framework ARTEX to attack hundreds of orgs — basedjensen · 2026-10-22