This paper investigates the use of inclusive language in EU and national administrative discourse and examines the extent to which Large Language Models (LLMs) converge with or diverge from human editing practices in the corresponding translation settings. Drawing on corpus-based translation studies and critical discourse analysis, the research is based on the compilation and analysis of parallel corpora of EU texts in English, French, and Greek, and comparable corpora of respective national texts (UK/France/Greece), exclusively including official administrative documents. The trilingual parallel EU corpora comprise 4,254,342 words across 1,770 files for the period 1993–2023: 1,282,845 English words, 1,570,898 French words, and 1,400,599 Greek words. The comparable national corpora comprise 3,573,199 words across 1,107 files: 1,256,469 words from the United Kingdom, 1,110,513 words from France, and 1,206,217 words from Greece. The LLM-assisted component examines three specific systems: GPT-5.3 Instant, Claude Sonnet 5, and Gemini 3.5 Flash.

As far as trilingual parallel EU corpora are concerned, the study focuses on the gender-oriented behaviour of LLM-assisted human revisions of EU translations in administrative contexts. The analysis demonstrates, on the basis of descriptive frequency comparisons, that LLMs systematically produce higher frequencies of inclusive language than human-edited reference corpora of EU texts. Rather than merely reproducing existing inclusive language strategies, the models appear to amplify them, revealing the tendency described as “algorithmic overcompliance”.

When it comes to comparable national corpora, the study shows that gender-oriented prompt engineering at inference time and subsequent human-guided editing can substantially improve convergence between LLM-assisted human-edited output and institutional practices. Prompt engineering here refers specifically to the provision and refinement of specialised gender-oriented instructions at inference time and does not involve model training or fine-tuning through parameter updates. Models guided by specialised prompts achieve higher descriptive levels of alignment, as measured through comparative frequency analysis, with exclusively human-edited administrative texts, although performance differences persist across languages, especially in morphologically complex contexts such as Greek. In particular, the findings indicate that the adoption of inclusive language remains uneven across languages, with English achieving the highest inclusive language usage, French showing moderate implementation, and Greek presenting comparatively lower levels. These results suggest that the diffusion of inclusive language practices is influenced not only by policy recommendations but also by language-specific morphological and institutional factors.

Overall, the findings raise important questions regarding the role of LLM-assisted human editing of EU multilingual and national intralingual translations in administrative communication, highlighting the need for language-dependent and context-sensitive approaches to the implementation of inclusive language in algorithmically mediated translation.