“Is This AI?”: A Danmaku-Located Phenomenology of How Audiences Meet AI-Generated Video on Bilibili

Danmaku provide public textual responses aligned to a video’s playback timeline. We examine danmaku on the eighteen prize-winning works of Bilibili’s first AI Video Creation Contest. We use within-work comment density to locate candidate moments for phenomenologically informed close reading, rather than to measure psychological absorption; five-second windows at or above each work’s 80th percentile define the primary sample, and a corpus inventory, sensitivity checks and a bilingual evidence log make the selection inspectable. In several selected intervals, appreciation of an image and explicit questions or remarks about its synthetic origin appear together: in a faux natural history documentary, viewers praise its familiar style and ask whether it is AI within the same five seconds. We call this juxtaposition doubled spectatorship and trace it through readings of astonishment at impossible images, uncanny likenesses, documentary texture and expressions of bodily response. The intervals also include technical fault-finding; the design neither shows simultaneous states in one viewer nor distinguishes AI video from CGI, animation or live-action video. The study contributes a documented sampling procedure and proposes, as a hypothesis for comparative and audience research, that origin-questioning in AI video may concern whether anything depicted was filmed at all.