Does your fact-checker work across languages and today's news?

Misinformation spreads faster than static benchmarks can measure. This team built CommunityFact—15,992 real claims across five languages and two domains—to test how well LLMs verify facts when they can search the web. Web access provided the biggest boost, yet models still picked sources humans rejected, a gap fixable through targeted retrieval strategies. The benchmark and Community Notes training data are released.