This research investigates how bias and toxicity mitigation techniques applied in English can transfer to other languages in multilingual large language models.

• LLMs show higher levels of harmful biases and toxicity when prompted in non-English languages
• Finetuning on specialized datasets in English can effectively reduce bias across multiple languages
• Different finetuning methods impact both bias reduction and the model's linguistic capabilities
• Critical for developing secure multilingual AI systems that maintain safety standards across all supported languages

For security professionals, this research provides valuable insights into creating safer multilingual language models that protect users from harmful outputs regardless of their language.

Original Paper: Cross-Lingual Transfer of Debiasing and Detoxification in Multilingual LLMs: An Extensive Investigation