Unverified50% confidenceOpinionExact time
The training data of large language models contains demographic biases, with English-speaking internet users and more highly educated groups being overrepresented, while the elderly, rural populations, and speakers of non-mainstream languages are underrepresented
1
Sources
50%
Confidence
Long-term
Relevance
9/28/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedEnglish content in mainstream large language model training corpora typically exceeds 50%, leading to significant performance gaps for non-English languages.71% similarUnverifiedNon-English speakers are at a systemic disadvantage in accessing information, participating in discussions, and building cross-cultural connections, forming a 'digital language divide.'71% similarUnverifiedGlobally, there are no more than a few hundred researchers with large-scale language model pretraining experience, making supply extremely scarce.71% similarUnverifiedEnglish data accounts for an overwhelming proportion of mainstream large language model training corpora; taking public datasets such as Common Crawl as an example, English content typically constitutes 40%–60% of the total, while languages such as Vietnamese and Chinese often account for less than 5%.70% similarVerifiedThe safety alignment of many models degrades significantly in non-English languages, and low-resource languages are especially easy to exploit because the human annotation data for RLHF alignment training is highly concentrated in English.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/957157API
curl https://kongchang.com/api/v1/knowledge/claims/957157MCP
get_claim(id=957157)