Beyond Automation: An Observability and Error Framework for AI-Enabled Marine Recreational Fisheries Monitoring

Monitoring marine recreational fishing remains difficult because effort and catches are spatially dispersed, participation is heterogeneous, and reporting is rarely mandatory. This narrative review asks what artificial intelligence (AI) can automate and when AI-derived observations can support defensible fisheries inference. We synthesise established on- and off-site monitoring methods, direct recreational-fisheries applications of AI, and transferable evidence from commercial fisheries and aquaculture. Across methods, the limitation is observability: each monitoring approach captures a subset of fishing, and greater data volume does not necessarily increase population representativeness. AI primarily expands the speed, scale and richness of information extraction within observable events while introducing its own measurement and model errors. We propose an observation-to-decision framework in which sampling and coverage error, sensor and measurement error, algorithmic error and estimation uncertainty are considered jointly. Transferability is treated as a domain-shift problem requiring external validation across sites, seasons, devices and species. The most defensible near-term architecture is a calibrated hybrid system in which probability-based surveys anchor population inference, direct observations provide validation, and AI-enabled cameras or angler images provide high-frequency measurement. The value of AI to fisheries management should therefore be judged by the validity of the complete monitoring system, not by algorithmic accuracy alone.