Dramatic translation presupposes more than rendering words and meanings from the source to the target language; it is a linguistic, intra- and intercultural continuous negotiation meant to preserve and/or transform the performative properties of the source text, creating an actable result adapted to the TL’s linguistic and cultural specificity.

Drawing on the findings of Hye K. Pae et al. (2026), the presentation discusses a contrastive analysis of human and LLM dramatic translations to see whether LLMs are able to translate dramatic texts dramatically, that is to preserve their innate theatricality, here understood as textual features (e.g. word pace, sentence rhythm, breathing pauses, cultural and extratextual elements such as word play or multi-layered meanings) that turn such texts from words on pages to words on stages, or whether there is still need for human intervention to either polish such translations or make them from scratch.

The analysis is multi-layered: on the one hand, it is meant (i) to contrastively analyse LLM and human dramatic translations from the perspective of their linguistic, cultural and theatricality preservation and/or adaptation, and (ii) to see whether LLM back-translation of both human and AI-generated translations yield the original texts and analyse the differences considering the same theatrical parameters.  

First, translated excerpts in Romanian, French and English (for non-English texts) from authors like Shakespeare or the Ancient Greeks were chosen trying to avoid important names or cues, possible indicators of the source texts. These already-used-in-theatrical-shows human translations, adapted to and adopted by the target language, are considered baseline translations as their extensive usage suggests that (at least) part of their dramatic features were transferred to the TLs. A comparison of theatricality features, such as those just mentioned, was also performed between the three languages.

Next, using different LLMs, such as Claude, the original source texts were translated into English (for non-English texts) and French, both widely-used languages where better quality is expected, and Romanian, a less-used language where more errors are anticipated. The AI and the human translations are compared based on the core features of dramatic texts, i.e, pragmatics (deictic anchoring and speech acts), semantics (subtext and puns), or syntax (rhythm and fragmented speech).

In order to see how far from the original these translations are and to test whether LLMs can recreate the original texts from human and AI-generated translations, these two types of translation were back-translated into the source language using different LLMs, taking into account that LLM training data likely include these famous authors. For the LLM translations a cross-model back-translation process was used in order to isolate model behaviours while preventing algorithmic self-validation.

H. K. Pae, D. Xia, H. Sun, et al. “Translation and Cross-Cultural Adaptation of the Chronic Rhinosinusitis Control Test for Global Use.” International Forum of Allergy & Rhinology (2026): . https://doi.org/10.1002/alr.70194

The scope of our analysis was further extended to a small sample of English untranslated dramatic texts that we translated into Romanian using LLMs and human experts (two linguists and two actors) to check them for the same pragmatic, semantic and syntactic criteria. The resulting outputs also underwent cross-model back-translation to measure their proximity to the original source texts.

The analysis is meant to see whether LLMs lack the creativity needed to preserve the theatricality of the source texts, but also to adapt them culturally and linguistically to the target language, features that are intrinsic to human translators,