Researchers have found that current safety alignment training for Large Language Models (LLMs) is heavily biased towards English. This can lead to stereotype-reinforcing outputs and propagation of harmful biases to non-English speaking communities. A new benchmark, called INCLUDE, has been introduced to address this cross-lingual gap. The benchmark evaluates LLMs against socio-cultural biases in six languages, including Hindi, Bengali, and Marathi. The results show that Bengali models have the highest bias scores, while English models have lower bias scores when open-sourced but higher bias scores when closed-sourced. --- Why it matters: This matters to AI researchers because it highlights a critical failure mode in LLMs deployed in multilingual environments. It also underscores the need for more inclusive and culturally sensitive training data to prevent the propagation of biases. Source: https://arxiv.org/abs/2608.18131