Large language models (LLMs) are rapidly becoming integrated into our daily lives. But are they truly multilingual, or do they speak other languages with an English "accent?" New research suggests the latter might be true, particularly for LLMs primarily trained on English data. These models, while capable of generating text in multiple languages, exhibit biases toward English vocabulary and sentence structures. Think of it like someone learning a new language—traces of their native tongue inevitably seep through. This research introduces novel metrics to measure this "accent" by comparing the vocabulary and grammatical structures of LLM-generated text to natively written text. The results show a significant gap, especially in languages like Chinese and French, which are typologically distant from English. For example, LLMs trained predominantly on English data tend to overuse passive voice constructions in Chinese, a structure less common than in English. Interestingly, even when generating in Chinese, the syntactic structures of some LLMs more closely resemble English than native Chinese. This "English accent" poses challenges for fair language representation and could disadvantage communities that speak lower-resource languages. However, the research also offers a solution: a method to "align" these models, improving their naturalness in other languages without sacrificing performance on general language understanding benchmarks. By fine-tuning models with datasets that contrast native text with synthetically generated "unnatural" text, researchers have demonstrated significant improvements in naturalness. This research highlights the growing importance of moving beyond simply evaluating LLMs on their ability to complete tasks and focusing on the nuances of how they generate language. It's not just about what they say, but *how* they say it. As LLMs become more prevalent, ensuring they speak other languages fluently—without an English accent—will be crucial for creating truly inclusive and equitable technology.