Politeness judgments draw on linguistic, non-verbal, and situational information, yet previous research has often examined these dimensions separately or reduced context to abstract social variables. This study compared GPT-4o with nine native Korean speakers in judging directive utterances from three Korean drama scenes where politeness predicted from utterance form and social variables differed from how the utterances were perceived in context. Participants watched each scene, rated the target utterance, and explained their judgment. GPT-4o received the same background information and scripts, together with utterance-level acoustic summaries and still images, and completed the task ten times per scene. GPT-4o produced highly consistent ratings, whereas human judgments varied widely. GPT-4o’s explanations addressed all eight predefined factors, interpreted non-verbal cues through fixed associations, and applied sociocultural norms as general rules. Human participants instead focused on a smaller number of decisive factors, inferred unstated context, drew on personal experience, and formed diverse narrative interpretations. These findings show that politeness is an interpersonal meaning constructed through the contextual interpretation of linguistic, non-verbal, and situational information. They further suggest that, to provide learners with appropriate feedback on politeness, LLMs need a more human-like capacity to weigh multiple factors flexibly in context.