The Influence of Prompt Language on Code Generation: A Comparative Analysis of Turkish and English

This study investigates the effect of prompt language (Turkish vs. English) on code generated by the Gemini 3 Pro model. Contrary to the common assumption that English yields superior results, our experiments reveal a significant performance difference: Turkish prompts achieved a functional correctness rate of 78.62%, decisively outperforming the 50% rate from English prompts on our unit test suite. However, this functional superiority did not equate to structural similarity. Analysis using Abstract Syntax Trees (AST) showed a low structural similarity of only 34.55% between the code pairs, indicating that the model's choice of algorithmic strategy was fundamentally altered by the input language. Furthermore, code generated from Turkish prompts was 28.22% more verbose, suggesting the model utilized more granular steps or a different syntactic approach. These findings demonstrate that the prompt language influences not only the model's comprehension of a task but also the fundamental logic it employs to generate a solution. This suggests the model does not simply translate the problem internally but accesses distinct problem-solving pathways based on the linguistic context of the prompt.