Artificial intelligence approaches, such as computer vision, are increasingly being explored for use in ecosystem monitoring. However aquatic systems and underwater video can pose a multitude of unique challenges, many of which are exemplified in video data collected at the giant Australian cuttlefish aggregation in South Australia. Cuttlefish are highly variable in size, shape and colour, both between and within individuals, and are adept at camouflaging among the complex macroalgal reef that forms their habitat. Building on the YOLO object detection model, we have developed a computer vision pipeline that can reliably detect cuttlefish to inform estimates of abundance and detect activity patterns. Despite its apparent simplicity, this detection task exposed challenges emblematic of the wider barriers to applying AI in underwater ecosystem monitoring. The model enables fast processing of past and present underwater video data from the region, allowing long-term monitoring of cuttlefish abundance and behaviour.