LLMs Infer Cultural Context but Fail to Apply It When Responding

Large language models can detect a user’s cultural background but routinely fail to apply that knowledge when generating responses, according to a study submitted in 2026 that introduces a new evaluation dataset called CAPRI [1]. The research, posted to arXiv on 16 June 2026, examines whether state-of-the-art LLMs adapt their answers to local conventions — specifically measurement units — after inferring a speaker’s cultural context [1]. The authors find that models “can infer cultural background and recall relevant conventions, but often fail to utilize the information to adapt their answers to the relevant cultural conventions, unless explicitly prompted to perform the tasks sequentially” [2]. The paper also evaluates how models handle time and quantity expressions, two dimensions of subjective language grounding that vary across cultures [1]. As cultural cues accumulate, the models do adjust their responses more, but their baseline assumptions are not neutral: the priors sometimes align with the model’s country of origin [2]. The work builds on earlier findings that LLMs overrepresent dominant cultures, particularly Western ones, while marginalizing others [1]. This pattern is consistent with broader observations in artificial intelligence research. Since the 2020s, generative AI has become widely available for producing text, images, and audio, and systems are often designed to generate human-like outputs [5]. The tendency of users to attribute human-like mental states to AI — known as AI anthropomorphism — can be especially powerful when outputs appear culturally fluent, even if the underlying model lacks genuine cultural understanding [4]. The CAPRI dataset, short for Cultural and Pragmatic Response Inference, provides a structured way to measure the gap between what a model knows about a culture and how it actually behaves in conversation [1]. The authors frame the resource as a tool for future research aimed at “narrowing the gap between cultural knowledge and culturally adaptive language generation” [2]. The study does not include direct quotes from the researchers, but the abstract makes clear that the failure to apply cultural information is systematic rather than occasional [1]. Machine learning, the broader field in which LLMs operate, relies on statistical algorithms that learn from data and generalize to unseen examples [3]. When training data skews toward certain cultural norms, the resulting models can inherit those biases. The CAPRI findings suggest that even when a model possesses the relevant cultural knowledge, the architecture does not automatically deploy it during generation — a limitation that may require explicit task decomposition or prompting strategies to overcome [1].