Generative AI tools, such as ChatGPT, are applied to a wide range of languagerelated tasks, including translation. Despite their current popularity among users and researchers and the impressive results obtained on several benchmarks (Kocmi et al., 2024a; Deutsch et al., 2025), their potential side-effects on languages and translations are still understudied (Vanmassenhove, 2025). The paradigm shift from Machine Translation (MT) to Generative AI Translation (GAIT) likely calls for a reconsideration of our assessment and evaluation metrics and practices. In this work, we focus on GAIT by analyzing translations from four multilingual large language models (MLLMs), mBART, Jamba-1.5-large, GPT 4o and DeepSeek R1 applied to three different domains (news, literature and poetry) for the English-Dutch language pair. Focusing on metrics related to lexical and textual diversity, we find that while GAIT text for literature if of significantly high lexical and grammatical richness, that is not the case for news and poetry. We also assess the homogeneity of AI-generated text through a set of clustering and classification experiments. In addition to a clear separation between human- and AI-generated content, our results indicate that GAIT output is more homogeneous among MLLMs.