Low-impact fishing gear, such as fish pots, could help reduce human’s impact on coastal marine ecosystems in fisheries but catch rates remain low and the harvest of resources used for baiting increases their environmental cost. Using black seabreams (Spondyliosoma cantharus) as target species in the Bay of Biscay, we developed and assessed the efficiency of biodegradable biopolymer-based baits (hereafter bio-baits) made of cockles (Cerastoderma edule) and different biopolymer concentrations. Through a suite of deep and machine learning models, we automatized both the tracking and behavior classification of seabreams based on quantitative metrics describing fish motion. The models were used to predict the interest behavior of seabream towards the bait over 127 h of video. All behavior predictions categorized as interested to the bait were validated, highlighting that bio-baits have a much weaker attractive power than natural bait yet with higher activity after 4 h once natural baits have been consumed. We also show that even with imperfect tracking models, fine behavioral information can be robustly extracted from video footage through classical machine learning methods, dramatically lifting the constraints related to monitoring fish behavior. This work therefore offers new perspectives both for the improvement of bio-baits and automatic fish behavior recognition.