Lost in Translation: Large Language Models in Non-English Content Analysis
Multilingual LLMs used for content moderation perform substantially worse in non-English (especially low-resource) languages; cautions against treating them as a fix for global moderation.
Executive summary
AI-generatedThis summary was generated by AI from the original report to make it easier to scan and cite. It is not a substitute for the source — read the original above.
This report from the Center for Democracy & Technology (CDT) examines the use of large language models for automated content moderation across languages, focusing on how performance differs between English and non-English text. The source page could not be retrieved directly for this summary, so the description here relies on the report's stated title and available characterization of its findings.
Based on that framing, the report finds that multilingual LLMs deployed for content-moderation tasks perform substantially worse when analyzing non-English content, with the gap most pronounced for low-resource languages that have less training data and fewer benchmark resources available.
The report's conclusion cautions platforms and policymakers against treating large language models as a ready substitute for human-driven or language-specific moderation systems in global content governance, arguing that uneven model performance across languages risks leaving non-English-speaking user populations with weaker protections against harmful content.
Think this summary is wrong? Contact us.