Quietly taking sides: The machine habitus and metaphorical framing in human and LLM-generated news

Large language models (LLMs) are increasingly used to draft the news. These systems may frame the news differently from human journalists. In particular, it remains less understood how they handle metaphor, a potent framing tool in news discourse. Following Bourdieu’s field theory, we compare LLMs’ metaphorical framing with the journalistic habitus of human reporting in two studies of German news on Chinese investment spanning a quarter century (with 2975 metaphors from human journalism and 1924 from LLM-generated news writing), and conduct a prompting experiment across four instructions and three recent LLMs (DeepSeek V4, GPT-5.5, and Claude Sonnet 4.5) with 4877 metaphors. We find that compared to journalists that favored deliberate, evaluatively charged figurative expressions, the LLM-generated news was conventional by default, framing investment as a procedural, depoliticized process. We regard this disposition as a “machine habitus”. It proved durable across prompted years and model versions, bound to a closed set of source domains, and generative , varying under instruction only within that set. Its conceptual sources and deliberateness moved sharply with the prompt, yet showed little variation across LLMs, while its evaluative orientation tracked the prompt, with only weak differences by model origin. Across two versions of LLMs, the disposition proved more durable than the volume of metaphor it governs: the newer LLMs wrote more figuratively yet remained conflict-averse. The study contributes to debates on automation, framing, and accountability in journalism.