In 2024, the Oxford Internet Institute studied misinformation spread across 81 countries.
Their finding: the most dangerous misinformation wasn't in English. It was in languages that English-language fact-checking tools couldn't read. WhatsApp forwards in Hindi. Facebook posts in Swahili. Telegram chains in Arabic. Viral claims in Tamil that never get fact-checked because the tools don't exist.
Here's the uncomfortable truth about the current state of NLP fact-checking:
95% of fact-checking models are English-only.
The LIAR dataset — the most cited benchmark in claim verification research — is entirely in English. FEVER, the gold standard for fact verification, is entirely in English. Most production fact-checking APIs? English only.
Meanwhile, India alone has 22 official languages and 500 million WhatsApp users. A false claim about a vaccine, an election, a riot — spreads in minutes in a language no existing model can verify.
This is not a model problem. It's an architecture problem.
Cross-lingual transfer learning has existed since 2019 — XLM-RoBERTa was pre-trained on 100 languages simultaneously. The capability is there. The application isn't.
Datasets like MM-COVID, CLEF CheckThat! 2023, and IndicGLUE exists precisely for this — multilingual misinformation benchmarks that almost nobody in the open-source community has seriously combined and trained on.
The gap between what's possible and what's been built is embarrassingly wide. Someone should close it.
This is exactly why I built Sift 🔍
Sift is an open-source multi-agent fact-checking pipeline — 5 agents, each playing a distinct role. But Sift today only speaks English. 🌐
And that's the problem I'm solving next. Someone should close this gap. I intend to. 🚀
🔗 Full technical breakdown of how Sift works: [(https://dev.to/ashg2099/i-built-an-open-source-multi-agent-fact-checker-heres-how-it-works-5eah)]
Top comments (0)