Lost – or Filtered – in Translation: Machine Translation, Large Language Models, and the Challenge of Hate Speech

Description

Twenty years after its launch in 2006, Google Translate stands as the dominant specialised Machine Translation (MT) tool worldwide (Yao 2026), enabling communication across languages and cultures. Nevertheless, modern MT approaches mirror biases rooted in training data (Savoldi et al. 2025), an issue that becomes particularly relevant when the content being translated is hate speech (HS) and slurs. The rise of genAI language tools has raised new questions about how potentially harmful content is managed across linguistic boundaries. The present study investigates how HS is handled in the translation task, comparing the behaviour of two well-established tools that differ significantly in their underlying design philosophy: the ‘vertical’ MT tool Google Translate and the same company’s ‘horizontal’ general-purpose large language model (LLM) Gemini. To do so, a curated set of subtitles (extracted from the parallel corpus OpenSubtitles2018) was fed to Google Translate (from English to Italian). In parallel, the same task was carried out on Gemini using a generic prompt, without prompt engineering (Boonstra 2025) or jailbreaking (Zhang et al. 2026) and adopting a qualitative longitudinal approach (Google Translate 2025/2026 vs. Gemini 1.5/2/3). The results show, on the one hand, that HS is translated by Google Translate without any mitigation, prioritising linguistic accuracy over content moderation; on the other, Gemini’s outputs changed over time: while early versions adopted specific translation strategies (e.g. deflecting or neutralising HS), the latest versions appear so sensitive to HS that the model refuses to complete the task altogether, as a result of its embedded ethical filters. We discuss the implications of this divergence for translation studies and the broader question of what we expect from AI language tools, linguistic accuracy or societal accountability, as well as what this still tells us about norms, expectations and cultural biases embedded in MT and LLMs algorithms.

REFERENCES:
Boonstra, L. (2025) Prompt Engineering. Google, https://www.kaggle.com/whitepaper-prompt-engineering (accessed: 29/04/26).
Savoldi, B., Bastings, J., Bentivogli, L., and Vanmassenhove, E. (2025). A decade of gender bias in machine translation. Patterns, 6(6).
Yao, R. (2026). Celebrating 20 years of Google Translate: Fun facts, tips and new features to try. Google, https://blog.google/products-and-platforms/products/translate/fun-facts-google-translate-20-years/ (accessed: 29/04/26).
Zhang, Z., Zhao, P., Ye, P. and Wang, H. (2026). Enhancing Jailbreak Attacks on LLMs via Persona Prompts. arXiv:2507.22171 https://doi.org/10.48550/arXiv.2507.22171

Period	3 Jul 2026
Event title	LLMs as Mirror, Colleague, Rival: 5th TSHD Digital Humanities Symposium, 2026
Event type	Conference
Location	Tilburg, NetherlandsShow on map
Degree of Recognition	International