LLMs Show Significant Cross-Lingual Safety Gaps, Especially in Indian Languages.

Namya Bhatnagar· August 20, 2026 View original

Key takeaways

  • LLM safety alignment is predominantly English-focused, creating vulnerabilities in other languages.
  • A new benchmark, INCLUDE, quantifies socio-cultural biases in Indian languages.
  • Open-source models showed highest bias in Bengali, while closed-source models showed highest bias in English.
  • Cross-lingual safety gaps can lead to stereotype reinforcement and harmful bias propagation.

Who benefits

AI DevelopmentGlobal TechSocial MediaCustomer ServiceEdTech

Summary

Current LLM safety alignments are heavily English-centric, leading to critical failures and biased outputs in non-English languages. A new benchmark, INCLUDE, reveals significant socio-cultural biases in LLMs across various Indian languages.

Research highlights a critical flaw in current Large Language Model safety protocols: their strong bias towards English-centric alignment. This English-first approach results in a substantial "safety gap" when LLMs are used in other languages, potentially propagating harmful stereotypes and biases, particularly in diverse linguistic environments like India. To address this, a new multilingual evaluation benchmark called INCLUDE has been developed. This benchmark, comprising 2,604 prompts in six languages including Hindi, Bengali, and Hinglish, was used to test ten different LLMs. The findings indicate that open-source models exhibited the highest average bias in Bengali, while closed-source models showed the highest bias in English, suggesting a complex and varied problem across different model types.

Why it matters

Professionals deploying LLMs globally must understand that English-centric safety measures are insufficient, risking reputational damage and ethical concerns due to biased outputs in non-English markets.

How to implement this in your domain

  1. 1Audit existing LLM deployments for non-English language performance and potential bias.
  2. 2Integrate multilingual and culturally specific safety benchmarks into LLM evaluation pipelines.
  3. 3Invest in diverse language data collection and annotation for more robust safety training.
  4. 4Collaborate with linguistic and cultural experts to develop nuanced safety guidelines for global markets.

Original post by Namya Bhatnagar

"arXiv:2608.18131v1 Announce Type: new Abstract: Current safety alignment training for Large Language Models (LLMs) are heavily English-centric. When such safety filters fail for non-English languages, the consequences are immediate and user-facing: voice assistants and spoken dia…"

View on X

Originally posted by Namya Bhatnagar on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses