Prompting in English: Not that Ideal After All
Given that English is one of the most commonly used languages across the globe with an estimated 1.5 billion speakers, one would expect it to be the clearest and most understood language to use on any generative AI platform...or so we thought. Recent research says, think again.
A study by the University of Maryland and Microsoft indicates that Polish ranks as the best language to use for prompting on most generative AI platforms. Using 26 languages with identical inputs, researchers tested how these languages performed on ChatGPT, DeepSeek, Gemini, Qwen and Llama. The outputs indicated that out of the 26 languages, Polish averaged an accuracy rate of 88 percent in completing given tasks, while French displayed an 87 percent accuracy level. This was quite interesting given that the amount of training data available for Polish to English is significantly less than what is available online for Chinese to English.
So, where exactly did the language of the majority of the world rank? Astoundingly, English ranked only sixth for this particular study, with an 83.9 percent accuracy average, just below Russian with 84 percent. The top ten ranking languages were:
- Polish
- French
- Italian
- Spanish
- Russian
- English
- Ukrainian
- Portuguese
- German
- Dutch
While this study is interesting and very recent (published just this past September), many AI-based community forums show that the bulk of users do prefer and recommend English as the go-to language for prompting. Additionally, most generative AI platforms have a built-in translation function which allows for inputs or results to be translated to a language of your preference.
In fact, another study reveals that translator prompts are actually revolutionizing global communication. A translator prompt is a specialized instruction designed to guide artificial intelligence systems to perform translation tasks with specific requirements for accuracy, cultural sensitivity and contextual appropriateness. One of the main challenges with machine translation is the understanding of context and cultural nuances that affect the translation output, which is where a human translator is still needed to oversee a machine-translated output. However, merging translator prompts with machine translation outputs has significantly improved translations by 15 percent.
An essential improvement in this space is the development of few-shot approaches for translator prompts. Instead of only relying on written instructions, prompts have examples included which show the AI platform the desired tone, style and context, so the LLM can then adjust its generated output accordingly.
So, what's the takeaway from all of this? Is English ideal or should we be prompting in a second language? The honest answer is simple – it's not like almost everyone is incredulously proficient in Polish, so while the Microsoft-UMD study may show the accuracy level of Polish to be quite high, this can be affected by a number of things like the quality of the input prompt, the type of task, the comparison of platforms chosen or the version of that model (for instance, it's not clear which GPT model was used for the study).
Given the ever-evolving nature of gen AI, language and the concept of prompting, the safe bet is to stick with the language you're comfortable with and instead of obsessing over whether it is correct or accurate enough, focus more on the strength of your prompt itself and what you're feeding the model to get a result that is most desirable. The next blog in this series will unpack the prompt engineering strategies to enhance prompts as researched by OpenAI themselves.
Posted by Ammaarah Mohamed on 12/15/2025