Speakers
Description
Recent advances in generative artificial intelligence (AI) have profoundly reshaped the landscape of translation. Large Language Models (LLMs) such OpenAI’s GPT series are now capable of producing fluent, coherent and context-aware translations that go beyond the word-for-word logic traditionally associated with earlier forms of Machine Translation (MT) systems (Kenny, 2022). Despite their apparent accuracy, however, these tools demonstrate uneven translation performance with texts belonging to different textual genres and degrees of specialization.
This paper investigates the potential and limits of current leading MT paradigms through the qualitative analysis of translations of texts situated along the continuum between general and specialized language (Cabré, 2007), with particular attention to semi-specialized discourse. The study draws on empirical data from a comparable corpus of travel literature and compares translations of French texts related to artistic and cultural heritage into English and Italian, produced by two Neural Machine Translation (NMT) systems (DeepL and Google Translate) and three generative AI models (ChatGPT-5.5, Claude Sonnet 4.6 and Mistral Medium 3.5).
The findings show that, while generative AI models often outperform traditional NMT systems in terms of fluency and overall textual coherence (Jiao et al., 2023; Hendy et al., 2023), they continue to display significant weaknesses in areas that are crucial for specialized and semi-specialized translation. These include inadequate handling of multi-word terms (e.g. tableau d’autel) and specialized lexical combinations (L’Homme, 1998, 2017) (e.g. colonne ronde), difficulties in semantic disambiguation (e.g. tombeau), and instances of so-called “hallucinations”. Such limitations can be attributed to the probabilistic, data-driven nature of LLMs, which lack genuine semantic understanding and are trained predominantly on large but largely generic and unevenly distributed corpora (Bender et al., 2021; Ciotti, 2023; Raus, 2024; Tekwa, 2025).
Against this backdrop, the paper revisits a question often considered outdated in the “Machine Translation Era” (Asscher, 2025, p. 4): are traditional dictionaries still worth using? The analysis strongly supports an affirmative answer. General bilingual and monolingual dictionaries remain essential for managing polysemy, stabilizing semantic correspondences, and preventing uncontrolled lexical variation – tasks that generative models perform implicitly and inconsistently (Mattioda, 2024). Even more crucial is the role of specialized lexicographic and terminological resources, which provide explicit conceptual structuring, domain-specific distinctions, and historically grounded knowledge that AI systems are currently unable to guarantee.
Rather than being rendered obsolete by generative AI, monolingual and plurilingual dictionaries emerge as complementary resources supporting human pre- and post-editing processes, guiding prompt design, and constituting high-quality supervised datasets for domain adaptation in neural and generative systems (Chu & Wang, 2018; Hansen, 2024). In this perspective, translation is increasingly configured as a hybrid activity in which human intelligence, lexicographic expertise, and AI interact complementarily rather than competitively.
The paper concludes that the future of translation does not lie in the replacement of dictionaries, but in their renewed centrality as linguistic, cognitive, and epistemological anchors within an evolving ecosystem of generative AI tools.