From online traces to upstream surveillance: Large language models uncover the hidden status of non-native freshwater apex predators in China

Abstract

Biological invasions are often recognised only after ecological damage has become visible in the wild, even though the human-mediated processes that drive introductions may already be underway through private rearing, trade, and release. This gap in upstream surveillance is especially consequential for non-native freshwater apex predators, whose impacts can propagate rapidly through recipient food webs. Here, we show that these human-mediated introduction dynamics can be reconstructed from online activities at the national scale. Using China as a test case, we developed a large language model-based pipeline to process and validate 366,210 posts on gars (Lepisosteidae) and snapping turtles (Chelydridae) from ten major online platforms spanning 2006 to 2024. We found that online rearing and trade were taxonomically uneven, platform-specific, and temporally dynamic, with already regulated species remaining dominant but several unregulated taxa also emerging repeatedly across rearing, trade, and sighting records. Wild sightings showed clear spatial and temporal structure and were associated with both socioeconomic and environmental context, indicating that an online trade-to-online sighting signal is detectable at the national scale. Together, these results show that dispersed online traces can be converted into usable ecological evidence for upstream surveillance of human-mediated introduction pathways. More broadly, our study provides a scalable framework for detecting online evidence of introduction pathways that conventional surveillance often misses.