The training-discovery paradox: Evidence from citizen science

Training enhances classification accuracy in citizen science by directing visual attention toward predefined features, but may constrain attention to unexpected objects or features. We examined this tradeoff through an analysis of 206 citizen science discoveries and through controlled experiments using eye-tracking and verbal protocols. The analysis of existing discoveries showed that untrained contributors produced a broader range of discovery types, including the majority of opportunistic and serendipitous discoveries. In the primary experimental study, explicitly trained observers showed 34% fewer fixations on non-diagnostic features compared to untrained observers. Training also reduced attention to secondary entities and constrained how observers described novel stimuli, a phenomenon we term classification transference. Trained observers described unfamiliar entities by referencing missing trained-category features rather than reporting new attributes. Implicitly trained observers preserved broader attentional sampling while maintaining sensitivity to variation. The second experimental study provided a robustness check with 84 additional participants using written responses and confirmed the patterns observed in the first experiment. These findings demonstrate that training shapes both where observers look and what they report, establishing a perceptual boundary on discovery potential that conventional accuracy measures cannot detect.