When Cow Urine Cures Constipation on YouTube: Limits of LLMs in Detecting Culture-specific Health Misinformation

A study presented at the 2026 ICWSM workshop reveals a critical blind spot in LLMs: they cannot reliably detect health misinformation when it's wrapped in culturally specific language. Researchers analyzed 30 multilingual YouTube transcripts promoting gomutra (cow urine) as a constipation cure—a claim that blends sacred Hindu traditions with pseudo-scientific terminology. They tested GPT-4o, Gemini 2.5 Pro, and DeepSeek-V3.1 with varying prompt tones and found that all three models systematically failed to flag the content as misinformation. The promotional videos used a rhetorical register that mirrors sophisticated debunking content, making it nearly invisible to models trained predominantly on Western, English-language corpora.

The problem goes deeper than simple language gaps. The researchers note that culturally embedded health misinformation doesn't look like ordinary misinformation—it uses sacred references, traditional authority, and pseudo-scientific jargon in ways that confuse LLMs. This cultural obfuscation also extends to gendered rhetoric and prompt design, compounding the analytical unreliability. The paper argues that cultural competency cannot be retrofitted through prompt engineering alone, calling for fundamentally different training approaches. For platforms like YouTube that increasingly rely on AI to moderate health content, this is a serious vulnerability, especially in Global South markets where such misinformation can have real health consequences.