“DeepL has observed that its specialized translation models are less prone to hallucination compared to general-purpose generative AI models when used for translation tasks.”