Skip to content

Multilingual Disinformation

Multilingual disinformation encompasses false and misleading content created or spread across multiple languages, as well as strategies for adapting disinformation narratives across linguistic and cultural contexts. The field addresses both detection (identifying false content regardless of language) and understanding how disinformation campaigns exploit language boundaries.

Key challenges

Language-specific variation: Disinformation campaigns adapt content to language-specific cultural contexts, framing narratives in ways that resonate locally. This requires detection models to account for language-specific linguistic markers and cultural references, rather than relying on universal feature sets.

Detection disparities: Machine-generated text detection systems show dramatic language variation—some languages (Polish, French) exhibit 4–5× higher rates of detected LLM-generated content than others (English, Arabic), suggesting either differential targeting or differential detector behavior.

Adversarial localization: Attackers may exploit language-specific gaps in content moderation. Non-English content receives less human review on social media platforms, creating an asymmetry in detection burden across languages.

Translation and paraphrasing: LLMs enable rapid content translation and paraphrasing, allowing disinformation to spread across language boundaries with minimal human effort. This complicates detection by creating large numbers of semantically equivalent but stylistically distinct variants.

Approaches

Multilingual model development: Fine-tuning pretrained models (mBERT, XLM-RoBERTa, multilingual LLMs) on balanced language distributions to avoid overfitting to high-resource languages (English, Spanish, French).

Cross-lingual transfer learning: Leveraging labeled data in high-resource languages to improve detection in low-resource languages through zero-shot or few-shot transfer.

Language-aware threshold calibration: Detectors trained on mixed-language data often exhibit language-dependent false positive/false negative tradeoffs; threshold adjustment per language improves reliability.

Language-specific corpora: Building multilingual disinformation datasets that cover diverse sources, platforms, and temporal windows to evaluate generalization across linguistic contexts.

Key papers